University of Liverpoollivrepository.liverpool.ac.uk/3005561/1/200867898_Jul16.pdf · Contents...
198
The retreat from overgeneralisation errors: A multiple-paradigm approach Thesis submitted in accordance with the requirements of the University of Liverpool for the degree of Doctor in Philosophy by Amy Bidgood. July 2016
Transcript of University of Liverpoollivrepository.liverpool.ac.uk/3005561/1/200867898_Jul16.pdf · Contents...
The retreat from overgeneralisation errors:
A multiple-paradigm approach
Thesis submitted in accordance with the requirements of the University of Liverpool
for the degree of Doctor in Philosophy by Amy Bidgood.
July 2016
Contents
Abstract
Chapter 1: Generativist and constructivist approaches to syntax acquisition 1
1.0 Introduction to the thesis .............................................................................. 1
1.1 General assumptions of generativist (nativist) approaches........................... 4
1.1.1 Principles and parameters ......................................................................... 5
1.1.2 Semantic bootstrapping ............................................................................. 7
1.2 General assumptions of constructivist approaches ....................................... 7
1.2.1 Early constructivist accounts .................................................................... 8
1.2.2 More recent constructivist approaches ...................................................... 10
1.3 Evidence for nativist and constructivist approaches ..................................... 12
1.4 Summary ....................................................................................................... 18
1.5 The roles of authors in papers ....................................................................... 19
Chapter 2: Argument structure overgeneralisation errors ............................. 21
2.1 Overgeneralisation errors ................................................................................ 21
2.2 Baker’s Paradox and the ‘no negative evidence’ problem .............................. 23
2.3 Theoretical approaches to explaining the retreat from overgeneralisation ..... 25
2.3.1 The semantic verb class hypothesis .......................................................... 25
2.3.2 The entrenchment hypothesis .................................................................... 29
2.3.3 The preemption hypothesis ....................................................................... 31
2.4 Integrating semantic and statistical approaches .............................................. 34
2.4.1 Distinguishing between entrenchment and preemption ............................ 42
2.5 Summary ......................................................................................................... 44
Chapter 3: The retreat from locative overgeneralisation errors: A novel
verb grammaticality judgment study ........................................................... 45
3.0 Fit within the thesis .................................................................................... 45
3.1 Introduction ................................................................................................ 46
3.1.1 The semantic verb class hypothesis ..................................................... 48
3.1.2 The frequency hypothesis .................................................................... 51
3.1.3 Existing evidence for the two accounts ................................................ 52
3.1.4 The locative alternation ............................................................................. 53
3.1.5 The present study ...................................................................................... 55
3.2 Method ............................................................................................................ 55
3.2.1 Ethics Statement ........................................................................................ 55
3.2.2 Participants ................................................................................................ 56
3.2.3 Design and materials ................................................................................. 56
3.2.3.1 Design ........................................................................................................ 56
3.2.3.2 Test sentences and animations .................................................................. 56
3.2.3.3 Novel verb training sentences and animations .......................................... 58
3.2.3.4 Grammaticality judgments ........................................................................ 59
3.2.4 Procedure ................................................................................................... 60
3.3 Results ............................................................................................................. 61
3.3.1 Preliminary analysis .................................................................................. 61
3.3.2 Testing the semantic verb class hypothesis ............................................... 63
3.3.3 Testing the frequency hypothesis .............................................................. 65
3.4 Discussion ....................................................................................................... 67
3.4.1 The role of semantics ................................................................................ 68
3.4.2 The role of frequency ................................................................................ 70
3.4.3 Explaining the retreat from overgeneralisation ......................................... 72
3.5 Acknowledgements ......................................................................................... 72
Chapter 4: How do children retreat from overgeneralisation error?
Evidence from the causative alternation ........................................................... 73
4.0 Fit within the thesis ......................................................................................... 73
4.1 Introduction ..................................................................................................... 74
4.1.1 Entrenchment ............................................................................................ 75
4.1.2 Preemption ................................................................................................ 76
4.1.3 Semantics .................................................................................................. 78
4.1.4 Recent evidence for the three accounts ..................................................... 79
4.1.5 The current study ....................................................................................... 82
4.2 Methods ........................................................................................................... 82
4.2.1 Frequency counts ....................................................................................... 83
4.2.2 Semantic ratings ........................................................................................ 84
4.2.2.1 Method ...................................................................................................... 84
4.2.2.1.1 Participants .......................................................................................... 85
4.2.2.1.2 Test items ............................................................................................ 85
4.2.2.1.3 Procedure............................................................................................. 86
4.2.3 Experiment 1 (adults – grammaticality judgments) .................................. 92
4.2.3.1 Method ...................................................................................................... 92
4.2.3.1.1 Participants .......................................................................................... 92
4.2.3.1.2 Test items ............................................................................................ 92
4.2.3.1.3 Procedure............................................................................................. 93
4.2.3.2 Results ....................................................................................................... 93
4.2.4 Experiment 2 (adults and children – grammaticality judgments) ............. 97
4.2.4.1 Method ...................................................................................................... 98
4.2.4.1.1 Participants .......................................................................................... 98
4.2.4.1.2 Test items ............................................................................................ 98
4.2.4.1.3 Procedure............................................................................................. 99
4.2.4.2 Results ....................................................................................................... 100
4.2.4.2.1 5- to 6-year-olds .................................................................................. 104
4.2.4.2.2 9- to 10-year-olds ................................................................................ 106
4.2.4.2.3 Adults .................................................................................................. 108
4.2.4.3 Discussion ................................................................................................. 111
4.2.5 Experiment 3 (children – production-priming) ......................................... 112
4.2.5.1 Method ...................................................................................................... 112
4.2.5.1.1 Participants .......................................................................................... 112
4.2.5.1.2 Test items ............................................................................................ 113
4.2.5.1.3 Materials .............................................................................................. 113
4.2.5.1.4 Procedure............................................................................................. 113
4.2.5.2 Results ....................................................................................................... 115
4.3 General discussion .......................................................................................... 119
4.3.1 Changes across development .................................................................... 123
4.3.2 How do children retreat from overgeneralisation errors? ......................... 123
4.3.3 Conclusion ................................................................................................ 124
Chapter 5: Children’s and adults’ passive syntax is semantically
constrained: Evidence from syntactic priming ................................................ 127
5.0 Fit within the thesis ......................................................................................... 127
5.1 Introduction ..................................................................................................... 128
5.1.1 The early semantic restriction account ...................................................... 129
5.1.2 The early abstraction account .................................................................... 130
5.1.3 A third possibility: The semantic construction prototype account ............ 131
5.1.4 The present study ...................................................................................... 132
5.2 Method ............................................................................................................ 134
5.2.1 Participants ................................................................................................ 134
5.2.2 Test items .................................................................................................. 134
5.2.3 Procedure ................................................................................................... 135
5.3 Results ............................................................................................................. 136
5.3.1 Adults ........................................................................................................ 141
5.3.2 Children ..................................................................................................... 143
5.3.3 Fine-grained semantic analysis ................................................................. 145
5.3.4 Summary ................................................................................................... 150
5.4 Discussion ....................................................................................................... 150
5.5 Acknowledgements ......................................................................................... 154
Chapter 6: Discussion ......................................................................................... 155
6.1 Introduction ..................................................................................................... 155
6.2 What do the studies in this thesis tell us about the retreat from
overgeneralisation? .......................................................................................... 155
6.2.1 Chapter 3: Locatives .................................................................................. 157
6.2.2 Chapter 4: Causatives ................................................................................ 159
6.2.3 Chapter 5: Passives .................................................................................... 161
6.3 Summary ......................................................................................................... 163
6.3.1 Evidence for the semantics hypothesis ...................................................... 163
6.3.2 Evidence for the entrenchment hypothesis ................................................ 165
6.3.3 Evidence for the preemption hypothesis ................................................... 166
6.3.4 Comparing frequency effects across studies ............................................. 168
6.3.5 The FIT account: An integrated approach ................................................. 169
6.4 Possible future research directions .................................................................. 173
6.5 Conclusion ....................................................................................................... 174
References ............................................................................................................ 177
Abstract
The retreat from overgeneralisation errors: A multiple-paradigm approach
Amy Bidgood
This thesis examines children’s argument structure overgeneralisation errors (e.g.
*Don’t giggle me!). Errors of this kind arise from children observing that certain
verbs can appear in more than one argument structure (e.g. The ball rolled/Homer
rolled the ball). This pattern can be usefully generalised to allow children who have
heard a verb produced in only one of these structures (e.g. The window opened) to
produce it in the other (e.g. Marge opened the window). The ability to generalise
patterns to new items is key to children becoming productive language users.
However, if they overgeneralise this pattern, errors will result: Bart giggled is
grammatical, but *Lisa giggled Bart (meaning Lisa made Bart giggle) is not.
This thesis tested three hypotheses designed to explain how children retreat
from such overgeneralisation errors, or, indeed, avoid making them altogether: the
semantic verb class hypothesis (Pinker, 1989); the entrenchment hypothesis (Braine
& Brooks, 1995); and the preemption hypothesis (Goldberg, 1995). Chapter 3 uses a
novel-verb grammaticality judgment paradigm to investigate overgeneralisation
errors in the locative construction (e.g. *Marge filled tea into the cup). Chapter 4
investigates overgeneralisation errors in the transitive and intransitive constructions,
using a grammaticality judgment paradigm with known verbs, as well as a
production priming paradigm designed to elicit errors from young children (e.g.
*Homer swam the fish). Finally, in order to investigate the role of semantics in
language development more generally, Chapter 5 moves beyond overgeneralisation
errors to investigate children’s acquisition of the passive construction (e.g. Bart was
helped by Lisa).
This thesis adds to a growing body of work demonstrating that none of the
individual theories (semantics, entrenchment, preemption) alone is able to explain
children’s retreat from overgeneralisation, and that an integrated approach, such as
that proposed by Ambridge and colleagues’ FIT account, is required to account for
the data. The thesis moves our understanding forward by demonstrating both that this
account can explain error patterns in production, and that the role of verb-in-
construction semantic compatibility (a key aspect of the FIT account) can explain
children’s acquisition of argument structure more widely.
1
Chapter 1: Generativist and constructivist approaches to syntax acquisition
1.0 Introduction to the thesis
This thesis investigates children’s argument structure overgeneralisation errors (e.g.
*Don’t giggle me!) and the mechanisms that have been proposed to explain how
children ‘retreat’ from such errors, or, indeed, avoid making them altogether.
Through a series of experiments, several accounts of mechanisms aiming to explain
the phenomenon are tested. However, it is first important to specify the theoretical
framework adopted throughout the thesis. This is not simply a matter of terminology,
but affects the assumptions about the knowledge children start out with, or develop
along the way, to help them solve the overgeneralisation problem.
Theoretical approaches to language acquisition may be broadly split into two
opposing views: generativist and constructivist. According to the generativist
account, children’s knowledge of syntax is abstract from the start of the process.
Under this approach, lexical items are the basic unit of language and children are
able to use their knowledge of syntactic rules to combine words (e.g. verbs and
nouns) into larger units (e.g. verb phrases). The constructivist account, in contrast,
assumes that constructions themselves (e.g. the transitive construction, X VERBed
Y) are basic units into which lexical items can be placed. Under this approach,
children are born with no knowledge of these constructions but, rather, they acquire
them from the input. Following naturally from this is the fact that many of children’s
early constructions appear to be lexically restricted, based on the lexical items the
child has heard in each construction in the input. This thesis investigates the
psychological reality of lexical effects in children’s acquisition of verb argument
structures through the examination of argument structure overgeneralisation errors
(e.g. *Don’t giggle me!).
This first chapter sets out in more detail the assumptions and predictions of
the generativist and constructivist approaches to syntax acquisition. Evidence for and
against the two approaches is then presented. While support is found for both
generativist and constructivist accounts, the balance of the evidence indicates that
lexical effects are a reality, and one that generativist accounts struggle to explain.
The constructivist approach therefore informs the studies presented in later chapters
of this thesis.
2
Chapter 2 examines in detail the main phenomenon of interest in this thesis:
argument structure overgeneralisation errors (e.g. *Don’t giggle me!). It begins by
explaining the origins of these errors and why the retreat from them may be such a
challenge for children (and for researchers attempting to explain how this might
come about). Three proposed mechanisms for the retreat from overgeneralisation are
then described: the semantic verb class hypothesis (Pinker, 1989), the entrenchment
hypothesis (Braine & Brooks, 1995) and the preemption hypothesis (Goldberg,
1995). Evidence in support of each is discussed, and the chapter concludes by
suggesting that elements of all three mechanisms may need to be combined in order
to successfully explain children’s retreat from argument structure overgeneralisation
errors.
Chapter 3 presents the first paper in this thesis (Bidgood, Ambridge, Pine &
Rowland, 2014). This paper reports a grammaticality judgment study investigating
overgeneralisation errors of the locative construction (e.g. Lisa sprayed the flowers
with water/Lisa sprayed water onto the flowers; c.f. *Homer poured the cup with
water/*Homer filled water into the cup). It therefore extends previous work (e.g.
Ambridge, Pine, Rowland & Young, 2008) to a new construction, with the locative
providing a critical test of the hypotheses under investigation because of the true bi-
directionality of the alternation and the fine-grained distinctions between the
semantic subclasses defined by Pinker (1989). The use of novel verbs is also a
particular strength of this paper (c.f. Ambridge, Pine & Rowland, 2012), as any
effects of semantics could not be attributed to participant’s previous experience of
the verbs in question appearing in locative sentences (i.e. any effect of semantics
must be independent of frequency effects). Results show effects of both verb
frequency and verb semantics, lending support to the entrenchment/preemption
hypotheses and the semantic verb class hypothesis, respectively.
Chapter 4 presents a multi-method paper investigating two of the most
frequent constructions in English: the transitive-causative (e.g. The man rolled the
ball) and the intransitive-inchoative (e.g. The ball rolled). This alternation is an
important test of the hypotheses due to its frequency (both full locative and full
dative sentences are rare) and the higher frequency of overgeneralisation errors
reported in these sentences types (see Pinker, 1989, pp. 22-25). So, whereas Bidgood
et al. (2014) provides a critical test of the semantics hypothesis in particular, this
paper investigates the ability of semantic and statistical approaches to explain the
3
errors children make most frequently. Extending the findings of previous work (e.g.
Ambridge, Pine, Rowland, Freudenthal & Chang, 2014) to the transitive-intransitive
alternation, using a large number of alternating and fixed-transitivity verbs in a
grammaticality judgment study, this paper took a different approach to semantics to
that of previous investigations of this alternation (e.g. Ambridge et al., 2008) by
viewing semantics on a continuum rather than in discrete classes. Taking a novel
methodological approach, syntactic priming was used to elicit overgeneralisation
errors from young children. Results from both methods show strong support for
entrenchment and semantic approaches, with more limited evidence for the
preemption hypothesis.
Chapters 3 and 4 provide evidence for frequency-based and semantic
accounts of children’s retreat from overgeneralisation errors. However, these
accounts must also be able to explain findings for constructions that do not involve
the production of errors if they are to be taken seriously as general mechanisms in
language acquisition. Chapter 5 therefore uses the priming method from Chapter 4 to
investigate the role of semantics in children’s acquisition of a construction known to
cause significant difficulties for young children in terms of comprehension: the
passive. This paper extends the findings of Messenger, Branigan, McLean and
Sorace (2012) to demonstrate that, while young children have abstract knowledge of
the passive construction (in line with generativist approaches), that knowledge is, in
fact, semantically constrained, in line with the theoretical approach developed
throughout this thesis.
Chapter 6 concludes the thesis, discussing the findings in Chapters 3, 4 and 5
and their implications for the entrenchment and preemption hypotheses and the
semantic account of the retreat from overgeneralisation errors. Overall, the findings
of the studies presented in this thesis all suggest the need for an integrated account
that can explain statistical and semantic effects, which will also be discussed in this
chapter.
Chapters 3, 4 and 5 are presented in a paper format rather than conventional
thesis chapter format. Chapter 3 has already been published (Bidgood et al., 2014)
and Chapters 4 and 5 are currently being prepared for submission to peer-reviewed
journals. The format of these chapters has been standardised so as to fit with the
thesis as a whole. For example, no abstract is provided and references are provided at
the end of the thesis. In order to be suitable for publication, the introduction for each
4
of the experimental chapters is relatively short in comparison with a conventional
thesis. More extensive introductory material is therefore presented in this chapter and
Chapter 2, with Chapter 6 providing discussion to draw the findings of the
experimental chapters together. Each experimental chapter is also introduced with an
explanation of how it fits within the thesis as a whole and how it links with the
preceding and following chapters. The primary reason for submitting the thesis in
paper format is so that the data can be more quickly and easily disseminated to the
wider academic community. The peer-review process has also been beneficial in
strengthening the paper presented in Chapter 3. Finally, I hope that, by publishing
work before the thesis is submitted, this will be beneficial for my future career.
The remainder of this chapter sets out the generativist and constructivist
approaches in detail. Evidence for and against each approach is then discussed, along
with why, ultimately, the constructivist approach was chosen as the theoretical
framework for the current thesis.
1.1 General assumptions of generativist (nativist) approaches
Generative approaches to adult grammar assume that sentences are ‘generated’ via a
set of formal rules in the domains of both morphology and syntax. For example,
combining a determiner (the) and a noun (bike) creates a noun phrase (NP; the bike)
(sometimes determiner phrase, although I will use noun phrase for consistency),
combing a verb (ride) with an NP creates a verb phrase (VP; ride the bike), etc.
Combining elements in the correct order by following the rules generates
grammatical sentences.
In principle, it is possible to have a generativist approach to language
acquisition that is non-nativist. However, generative approaches often assume that at
least some aspects of language must be innate. At the core of this argument is the
complexity of the task facing young children in acquiring their native language(s):
the input they receive from their environment could not possibly provide them with
sufficient evidence to learn all of the rules, or restrict them appropriately. This
argument is known as the ‘poverty of the stimulus’ (Chomsky, 1980): there are
simply too many possible rules children might posit in trying to work out which ones
apply to their language, despite hearing millions of words and sentences, if they did
not start out with a set of basic assumptions.
5
1.1.1 Principles and parameters
Famously, Chomsky (e.g. 1981) argued for a Universal Grammar (UG). Unlike
descriptive grammars of individual languages, UG was designed to highlight the
features of grammar that relate to all languages. These universal features give
children a starting point: there are features they know to be true (‘principles’); others
they know might be possible in their specific language (‘parameters’); and things
they know are not possible (‘constraints’). All of these enable children to correctly
interpret complex sentences, avoid grammatical errors, and become fluent speakers
of their native language(s) within just a few years.
One universal principle relates to syntactic categories: all languages have
certain syntactic categories, such as NOUN (although others, such as
DETERMINER, are not universal). Baker (2003) suggests that children have innate
knowledge that labels for objects are members of the NOUN category. Children are
thus able to populate their NOUN category with the object labels they hear – baby,
teddy, light, etc. Knowing which words belong to which categories helps children to
parse speech and to set language-specific parameters. This, in turn, helps them to
quickly assimilate new examples into the NOUN category, including abstract nouns,
through the use of distributional regularities in the input (e.g. Valian, Solt & Stewart,
2009).
An example of a parameter, which varies across languages, is that of head
direction. The ‘head’ of a phrase is usually what gives it its name, so a verb is the
head of a VP, a preposition is the head of a prepositional phrase (PP), etc. English is
a head-initial language: the verb comes at the start of the VP [Monkeys] eat bananas;
the preposition comes at the start of the PP in the garden. In contrast, Hindi is a
head-final language: the verb comes at the end of the VP [Bandarom] kēlē khānē
[monkeys bananas eat]; the preposition (actually a postposition here) comes at the
end of the PP bagīcē mēm [garden in]. As head direction varies across languages,
children must use language input to work out whether to set this parameter to head-
initial (e.g. for English) or head-final (e.g. for Hindi). For a summary of several other
proposed principles and parameters, see Ambridge and Lieven (2011: 122-123).
Unlike principles and parameters, which set out what is certain or possible in
a language’s grammar, constraints tell a child what is not possible (Crain &
Thornton, 2012). Constraints therefore help children to avoid producing grammatical
6
errors in their own speech and to avoid misinterpretation of complex speech in the
input. An example of a constraint is that of structure dependence (e.g. Crain &
Nakayama, 1987), which ensures that children posit rules based on the abstract
internal structure of a sentence, rather than on the surface order of words. The
operation of this constraint is often exemplified by the formation of complex yes/no
questions in English. To change a simple declarative sentence containing the
auxiliary be (1) into a yes/no question (2), the following rule is sufficient: invert the
auxiliary verb and the subject NP.
(1) All the monkeys are eating bananas.
(2) Are all the monkeys ___ eating bananas?
However, to change a complex declarative sentence (3) containing a restrictive
relative clause (italicised) into a grammatical yes/no question (4), a more specific
rule is required. The declarative sentence contains two auxiliary verbs from which to
choose, and moving the wrong one would result in an ungrammatical utterance (5).
(3) All the monkeys who are playing are eating bananas.
(4) Are all the monkeys who are playing ___ eating bananas?
(5) *Are all the monkeys who ___ playing are eating bananas?
The error in (5) would result if a child posited what seems to be the simplest rule:
invert the first auxiliary verb and the subject NP. However, thanks to the structure
dependence constraint, children will never posit this incorrect rule as they will
always take the abstract internal structure of the declarative sentence into account.
The correct rule is therefore posited: invert the auxiliary verb of the main clause and
the subject NP.
Under a generative approach, then, children acquire a set of rules, based on
the innate principles and constraints of UG, and the parameters they have set through
linguistic input. Because of this, children are able to rapidly acquire the grammar of
their language and avoid errors in their speech.
7
1.1.2 Semantic bootstrapping
Pinker’s (1989) version of semantic bootstrapping (see also Pinker, 1984) suggests
that innate linking rules complement other forms of innate knowledge to help
children begin the process of acquiring their language’s grammar. Through UG,
children are born knowing about syntactic categories (N, V), phrases (NP, VP) and
syntactic roles (subject and object). In addition, children have innate knowledge of
semantic (or thematic) roles, such as AGENT and PATIENT. The semantic
bootstrapping hypothesis proposes that children also possess innate linking rules
allowing them to map semantic roles onto syntactic roles, thus facilitating the
acquisition of argument structure. For example, the agent of a causal action (e.g. The
dog in The dog chased the cat) maps onto the subject role, while the patient (e.g. the
cat) maps onto the object role. As it is possible to observe from the environment
which is the agent and which is the patient and, assuming that they have acquired the
lexical items dog and cat, children are able to link the dog to the syntactic role of
subject and the cat to the syntactic role of object. This gives children the information
they need to work out that English word order is subject-verb-object (SVO).
Semantic bootstrapping is proposed as a mechanism that enables children to
break into UG and start putting in place the grammar of their language. Once
children have worked out that English has SVO order, for example, they will begin
to comprehend less concrete examples whose interpretation is not immediately
obvious from the environment. Indeed, children are able to determine syntactic roles
in sentences even when they have no idea what the sentences means, as in the
ubiquitous example, The situation justified the measures. Semantic bootstrapping,
then, allows children to break into their innate knowledge and use it to build the
grammar of their language. Once this grammar is built, they are able to use
distributional analysis to work that, for example, situation must be a noun.
1.2 General assumptions of constructivist approaches
In contrast to nativist approaches, constructivist accounts of language acquisition
assume no innate, language-specific knowledge. Instead, these approaches suggest
that general cognitive systems, such as categorisation, enable children to construct a
grammar from the input they receive. Children are not a ‘blank slate’, but they do not
8
require language-specific mechanisms to be encoded in their genes in order to be
able to learn language: the input they receive is argued to be perfectly sufficient for
children to succeed in acquiring their native language (i.e. there is no ‘poverty of the
stimulus’). Features of the input are also key in explaining the patterns of errors in
children’s language production that prove troublesome for nativist approaches.
Rather than building sentences by following a series of rules, as in generative
grammar, constructivist approaches posit a series of ‘constructions’ into which
lexical items can be placed to form sentences. For example, the construction for a
simple transitive-causative sentence might be something like N1 V N2, where N1
acts on N2, causing N2 to be affected in some way. So, in the sentence Bob annoys
Wendy, Bob (N1) acts on Wendy (N2) causing her to become annoyed. The
approaches outlined below illustrate different theories of how children might acquire
a construction grammar. Importantly, though, none of these theories rely on
underlying, innate grammatical knowledge to account for the data.
1.2.1 Early constructivist accounts
One of the earliest instantiations of a construction grammar is Braine’s (1963) ‘pivot
grammar’. Through examination of children’s earliest 2-word utterances (when they
are at the very beginning of syntax development), Braine noticed that these
utterances tend to be fairly limited for the first few months, before increasing
exponentially from 5 or so months after the first combinations appeared. This ‘first
phase’ of multi-word speech seemed to be characterised by a limited number of
words that always occurred in first or second position, and a greater variety of words
with which they combined, e.g. see boy, see sock, see hot; byebye plane, byebye man,
byebye hot; boot off, light off, water off. Braine termed these frequently-used words
‘pivots’, onto which more flexible ‘X-words’ could attach. X-words were essentially
all words in the child’s vocabulary except for the pivots, and could thus appear (in
principle) with any pivot word. Hence, children were provided with a simple way to
construct a number of combinations of the types pivot + X or X + pivot.
However, Brown (1973) argued that the three children studied by Braine
appear to constitute the only evidence for pivot grammars, with children in other
studies not conforming to these patterns (although he did note that the other children
discussed were at a more advanced stage of development). Firstly, what constitutes a
9
pivot varies widely from child to child, and even within children (verbs, pronouns,
prepositions, adjectives…). Secondly, one of the defining characteristics of a pivot
word is that it only occurs in one position (first or second), but this did not appear to
be the case for all children (e.g. Bowerman’s 1973 cross-linguistic study). Pivot
words also appear on their own (i.e. not in combination with an X-word) and some
combinations consist of two X-words or even two pivots. A pivot grammar strategy
cannot, therefore, be a universal strategy used to acquire syntax. One final problem
facing Braine’s pivot grammar is that, even if children do start out with a series of
pivot + X schemas, it is not clear how they would be able to move from this to more
advanced stages of syntactic development. The idea of pivot grammar has therefore
been built upon by Braine (1976), Bowerman (1976) and Maratsos and Chalkey
(1980), amongst others, to create accounts of development.
Further developing the idea that children’s initial syntactic knowledge
develops from lexically specific schemas, Tomasello (1992) outlined the Verb Island
Hypothesis. Tomasello argued that the sort of fixed word-order expressions
characteristic of pivot grammar do not reflect syntactic knowledge at all (cf. Ninio,
2014). Rather, the words simply appear in the order that children have heard them in
the input. Moving from this stage to abstract syntactic knowledge is a complex task.
Children begin with lexically specific knowledge: “in English, when you say eat, you
first say the person who’s doing the eating, then eat itself, then the thing being
eaten”. This leads to a schema such as [eater eat eatee]. The child builds a variety of
lexically-specific schemas: [chaser chase chasee], [kicker kick kickee], [jumper
jump], [talker talk], etc. By extrapolating across these schemas, something more
abstract results: the first person mentioned in all of these schemas is the one initiating
the action, i.e. the agent. Similarly, when a second person/object is given, they are
often affected by the action; noticing this allows children to create a patient category.
Words that come between the agent and patient are the action words (i.e. the verbs).
(Other information can also help to form categories: words ending in –ing can also
help form a verb category, for example. Morphology is especially important for
languages with case marking, and vital if word order is free.) Eventually,
extrapolating from verb-specific schemas allows for the creation of completely
abstract constructions of the N1 V N2 type.
10
1.2.2 More recent constructivist approaches
More recent constructivist approaches (e.g. Tomasello, 2003) are based on the
theories outlined above, although they do not posit such a reliance on verbs.
Nevertheless, at the core of these approaches is the theory that children’s syntactic
knowledge is built on lexically specific items (fixed phrases such as I want it or I’m
doing it), which are generalised to schemas (such as I want X or I’m X-ing it). The
importance of frequency in the input is also more firmly established, explaining why
certain constructions are more likely to be learnt early and also why lexical effects
are so often observed in children’s early language development.
Evidence from corpus and diary studies supports the view that children’s
earliest multi-word utterances are lexically constrained (e.g. Dąbrowska & Lieven,
2005; Lieven, Pine & Baldwin, 1997; Pine & Lieven, 1993; see e.g. MacWhinney,
1975, for cross-linguistic evidence). For example, Pine, Lieven and Rowland (1998)
studied the early combinations of 12 children, finding a lack of overlap in the main
verbs used with different auxiliaries and the nouns used as subjects and objects in
transitive sentences. These effects did not appear to reflect the input directly,
however, with children producing a disproportionate number of sentences with I in
subject position compared to child-directed speech. Lieven, Behrens, Speares and
Tomasello (2003) analysed the multi-word speech of a single 2-year-old child. Of the
utterances the child had not produced before (which accounted for only 37% of the
data), the majority required only a single change from a previously-produced
utterance. Many of these changes involved the substitution or addition of a noun.
Thus, many of this child’s utterances appeared to be based on fixed phrases or simple
schemas of the I want X type.
Evidence from experimental studies also provides support for lexical effects
in language development. In Childers and Tomasello’s (2002) elicited production
study, children aged 2;4-2;10 underwent training with real verbs in SVO sentences,
either containing pronouns or full noun phrases (e.g. He’s pulling it; The cow’s
pulling the car). At test, participants were required to use a novel verb to describe a
new scene, but only those who had received training including pronouns were able to
do so (e.g. He’s meeking it). This finding demonstrates that, although young children
have knowledge of the SVO construction, their ability to use it is affected by the
11
frequency of the frame in which the verb occurs (although see Fisher, 2002, for an
alternative explanation).
Evidence from additional experimental studies suggests that fixed phrases
and schema play a role in language development in other ways. Bannard and
Matthews (2008) tested 2- and 3-year-old children’s ability to repeat phrases, such as
sit in your chair, which is heard frequently in the input as a four-word phrase, and sit
in your truck, which is not. They found that children were significantly more
accurate at repeating the frequently heard sequences than the less familiar ones.
Following up on this, Matthews and Bannard (2010) used a similar task to test the
reality of slot-and-frame schemas in children’s language. They chose phrases such as
a piece of X: frequently heard three-word phrases (the ‘frame’) that varied in terms of
the final word (the ‘slot’). Children were better able to repeat four-word phrases
when the words that fill the slot are more variable in the input and, thus, led to
children having a more abstract schema. Together, these studies support the view that
children store phrases such as sit in your chair as whole units, but that they also store
phrases such as a piece of X as slot-and-frame schemas.
Through the processes of analogy and distributional analysis, children build
on lexically specific knowledge to create fully abstract constructions, such as the
transitive N1 V N2 construction. Showing support for the psychological reality of
analogy and distributional analysis, evidence from artificial language learning studies
has demonstrated that infants are able to learn simple ‘grammars’ through
distributional analysis, and that they are able to generalise these to new instances (see
Gómez & Gerken, 2000, for a review; for similar findings with adults, see e.g.
Altman, Dienes & Goode, 1995; Reber, 1969; 1989). Gómez and Gerken (1999)
trained 11- and 12-month-olds on an artificial grammar, using the head-turn
preference procedure. In the training phase, infants listened to strings of ‘words’
which conformed to the grammar. At test, infants listened to new strings that either
did or did not conform to that same grammar. Infants listened significantly longer to
the strings which conformed to the grammar they had just learned than to the
‘ungrammatical’ strings. This demonstrated that, with just a few minutes of
exposure, children were able to learn the rules of a very simple grammar and
generalise these rules to new examples.
12
Before reaching an adult-like stage of syntactic abstraction, children may
create a number of different ‘transitive’ schemas. Ambridge and Lieven (2015)
suggest the following examples (although they stress that these have not been tested):
Contact (non-causative) [AGENT] [ACTION] [PATIENT]
(John hit Bill)
Causative [CAUSER] [ACTION] [CHANGE]
(John broke the plate)
Experiencer-Theme [EXPERIENCER] [EXPERIENCE] [THEME]
(John heard Bill)
Theme-Experiencer [THEME] [EXPERIENCE] [EXPERIENCER]
(John scared Bill)
“Weigh” Construction [THING] [MEASURE/COST/WEIGH]
[AMOUNT]
(John weighed 100lbs)
“Contain” Construction [CONTAINER] [CONTAIN] [CONTENTS]
(The tent sleeps four people)
The examples at the top of this list (contact and causative) are the most frequent, and
therefore prototypical, examples of the transitive construction, whereas those at the
bottom are the least prototypical (for discussion of prototypes in language, see
Ibbotson & Tomasello, 2009).
1.3 Evidence for nativist and constructivist approaches
Innate knowledge of language might seem a reasonable way to explain children’s
rapid, and relatively error-free, acquisition of their native language. However, this
approach struggles to explain some features of children’s language. Logically, once a
rule has been acquired, the child should never make a mistake with that grammatical
structure again. However, the following evidence suggests that this is not the case:
error rates tend to vary across different lexical items, and that this variation is not
random. Note that some generativist theories (e.g. Head-Driven Phrase Structure
Grammar, Pollard & Sag [1994], Lexical-Functional Grammar, Bresnan [2001]) are
better able to deal with lexical effects than those which assume a full dissociation
13
between syntax and the lexicon (e.g. minimalism, Chomsky [1993], X-bar theory,
Jackendoff [1977]).
In an experimental study, Kidd, Lieven and Tomasello (2006) asked children
aged 2;10 to 5;9 to repeat sentences containing sentential complements (e.g. I think
she is riding away on a horse). Half of the sentences were grammatical and half
ungrammatical (e.g. *I think him running away from the dog). The complement-
taking verbs were either high-frequency (e.g. think) or low-frequency (e.g. pretend)
in corpora of child-directed speech. Results showed that children were better able
both to repeat grammatical sentences and to correct ungrammatical sentences
containing high-frequency than low-frequency complement-taking verbs. These
findings suggest that frequency information plays a role in language acquisition,
something that is not predicted under nativist accounts of language acquisition.
Similar lexical effects have been observed, in both corpus and experimental studies,
with other modal and auxiliary verbs (Pine et al., 1998; Rowland & Theakston, 2009;
Theakston, Lieven, Pine and Rowland, 2005), negation (Cameron-Faulkner, Lieven
& Theakston, 2007), and inflection (Wilson, 2003).
Further evidence of lexical effects have been demonstrated at sentence level
through the use of ‘weird word-order’ studies, with both novel and known verbs
(Akhtar, 1999). English has SVO word order (cf. Hindi, above, which has SOV word
order). In weird word-order studies (in English), the experimenter describes an event
(e.g. a doll pushing a toy car) using an ungrammatical sentence in which the word
order has been changed from SVO (Dolly pushed the car) to SOV (*Dolly the car
pushed). Children are then encouraged to describe a version of the same event. The
phenomenon of interest is whether children will imitate the experimenter’s
ungrammatical word order, or ‘correct’ it, by reverting to SVO. Matthews, Lieven,
Theakston and Tomasello (2005; see also Abbot-Smith, Lieven & Tomasello, 2001)
tested children aged 2;9 and 3;9 with English verbs of high, medium and low
frequency in the input (e.g. push, shove and ram, respectively). They found that 2-
year-olds were more likely to imitate the ungrammatical word order with low-
frequency verbs than medium- or high-frequency verbs. 3-year-olds, on the other
hand, were more likely to revert to the grammatical SVO order than to imitate the
ungrammatical word order with verbs of any frequency. Not only has this study
demonstrated lexical frequency effects, but it provides evidence of much more
gradual acquisition of syntax than an innate principles-and-parameters approach
14
would allow for (as the head direction parameter would either have been set or not,
and could not be in an intermediate state).
The bootstrapping proposals were designed to counter another issue for the
UG account, namely how children are able to use this knowledge to analyse the
strings of sounds they hear in their input. However, these proposals themselves are
not without their own problems. For example, semantic bootstrapping appears to
work quite well for canonical word orders in English (and children rarely hear
examples of the passive, for example, in which the order of agent and patient is
reversed: The dog chased the cat/The cat was chased by the dog). However, so-called
ergative languages do prove problematic for Pinker’s (1989) proposal. Nominative-
accusative languages like English treat the subjects of both transitive and intransitive
sentences in the same way, with the objects of transitive sentences being treated
differently. This can be seen in the case-marking system (note that case is only
explicit on pronouns in English): the subjects of both transitive and intransitive
sentences have nominative case (e.g. He chased Lisa; She ran), whereas the objects
of transitive sentences have accusative case (e.g. Bart chased her). In contrast, in
ergative-absolutive languages, such as Basque, the subjects of transitive sentences
are treated in one way (they have ergative case), whereas the subjects of intransitive
sentences are treated in the same way as the objects of transitive sentences (they have
absolutive case). In the following example, note the different case markings on the
first person singular pronoun and the determiner (adapted from Ezeizabarrena and
Larrañaga, 1996: 959):
(6) Nik lagun bat ikusi dot
I[erg] friend one[abs] seen has
‘I have seen a friend’
(7) Lagun bat etorri da
Friend one[abs] come has
‘A friend has come’
(8) Lagun batek ni ikusi nau
Friend one[erg] I[abs] seen has
‘A friend has seen me’
15
If children always map AGENT/ACTOR to the subject role, they might
assume that all initiators of the action/event, i.e. those in the subject position, would
have the same case marking. In this sense, ergative languages may be problematic for
the linking rules proposed in Pinker’s (1989) semantic bootstrapping account. The
proposed solution to this is that, rather than marking participants with the same
syntactic roles (as in accusative language), the morphology of ergative languages
mark the similarity of semantic roles. For example, the object of the verb see is a
theme (Lisa saw him), as is the subject of the verb laugh (He laughed); in ergative
languages, the pronouns in both of these sentences would have the same (absolutive)
case markings. However, the real problem for the semantic bootstrapping hypothesis
comes from ‘split-ergative’ languages (for a brief summary, see Pye, 1990: 1294).
These languages sometimes act like accusative languages and sometimes like
ergative languages. One example of a split-ergative language is Dyirbal, an
Australian Aboriginal language, which behaves like an ergative language with nouns
and third person pronouns, but like an accusative language with first and second
person pronouns (Dixon, 1979: 63). It seems impossible for innate linking rules to be
able to successfully explain children’s acquisition of these languages (or, indeed,
how such languages would have evolved in the first place, if innate linking rules link
syntactic and semantic roles consistently).
Constructivist approaches are, of course, not without their critics. Arguing
against the assumption that children’s first combinations primarily consist of rote-
learned, fixed phrases, Ninio (2014) presents an analysis suggesting that early 2-
word combinations are, in fact, syntactic phrases resulting from children’s use of
productive rules. Ninio analyses the telegraphic speech of young children (e.g. want
bottle, bring chair), focussing on their ungrammatical use of bare nouns (i.e. nouns
which, in adult speech, would require a determiner in order to be considered
grammatical; want [a] bottle, bring [the] chair, etc.). Analysis of child directed
speech shows that adults do not make this kind of error when addressing children
(although they do produce a number of grammatical, 2-word verb + noun utterances,
such as with proper and plural nouns), so the errors in children’s speech cannot be
due to imitation. At the same age, these children produce large number of determiner
+ noun 2-word utterances, so errors also cannot be explained by children simply not
being able to use determiners. Ninio suggests that children are using a productive
rule, of verb + single-word object to create these telegraphic utterances, which she
16
posits is derived from hearing grammatical verb + noun and verb + pronoun
utterances in the input. Ninio argues that these findings are best explained by
generativist accounts and, indeed, they seem to count against the fixed-phrase stage
of acquisition suggested by constructivist approaches. She suggests that performance
limitations also play a role. It should be noted, however, that these findings do not
necessarily count against a more abstract schema, such as action + thing, which
might be formed under a constructivist account. (The lack of specificity in
constructivist accounts in terms of which constructions/schemas children form, is
also problematic for current versions of this theoretical approach.)
Inherent to constructivist approaches to language acquisition is the
assumption that young children’s knowledge of syntax is lexically restricted and,
therefore, not fully abstract. These approaches therefore face challenges from studies
suggesting that, in fact, even very young children have abstract syntactic knowledge.
Gertner, Fisher and Eisengart (2006) used an intermodal preferential looking
paradigm (IPLP) to test if children were able to use word order to determine the
visual scene to which a transitive sentence containing a novel verb was referring.
Children aged 21 and 25 months old watched videos in which a bunny and a duck
performed novel actions on each other (e.g. in picture one, the bunny was pulling the
duck along by its legs while the duck lay in a wagon; in picture two, the duck tipped
the duck, who was sitting in a rocking chair). The children then heard the sentence
The duck is gorping the bunny. Even the youngest children looked significantly
longer at the correct picture, indicating that they were able to use abstract syntactic
knowledge to understand that The duck is gorping the bunny meant that the duck is
doing something to the bunny, and not vice versa.
While Gertner et al.’s (2006) study demonstrated that very young children are
indeed able to use some sort of abstract syntactic knowledge to interpret sentences
containing novel verbs, other studies using similar paradigms have shown that this
knowledge is nevertheless restricted. Chan, Meints, Lieven and Tomasello (2010)
used a similar paradigm to Gertner et al. (2006), except that the two videos showed
identical actions, just with the agent and patient roles reversed. Under these
conditions, 24-month-old children were unable to select to the correct interpretation,
and it was not until they were 33 months olds that children performed above chance
on this task. Noble, Rowland and Pine (2011) replicated the findings of Gertner et al.
(2006) with novel transitive verbs using a forced choice pointing paradigm with
17
children aged 27, 31, 40 and 51 months old (note that all of these children were older
than those in Gertner et al.’s 2006 study). This paradigm is similar to IPLP except
that children have to make a conscious choice to point at the scene that they think
matches the sentence. However, children did not perform above chance in a
conjoined agent intransitive condition (e.g. The duck and the bunny are daxing) until
after the age of three years. Finally, Fisher (1996) tested children’s ability to
comprehend giver and receiver roles. Watching a single video of two female
participants exchanging an object, the children heard a sentence such as She’s
trasking the balloon to/from her. They were then asked to Point to the one who was
trasking. Children were able to do this in the canonical to condition, but not the more
unusual from condition, even when they were as old as 5 years.
Taken together, the IPLP and pointing tasks described here indicate that,
while children as young as 21 months old demonstrate some abstract knowledge of
canonical transitive constructions, knowledge of verb argument structure continues
to develop over the third year of life. With more complex, non-canonical ditransitive
events, this knowledge is not yet adult-like at the age of 5 years. Continuing
development of argument structures is compatible with a constructivist view of
language acquisition, but not so easy to explain under a generativist-nativist view.
Syntactic priming studies have allowed very young children to demonstrate
their abstract syntactic knowledge in production. Syntactic priming (or structural
priming) refers to the phenomenon that children and adults tend to re-use syntactic
constructions that they have just heard (for reviews, see Branigan, 2007, and
Pickering & Ferreira, 2008). Of particular relevance to the question of abstractness of
syntactic knowledge is the assumption that priming is only possible when the person
being primed has an abstract syntactic representation of the sentence structure in
question (Branigan, Pickering, Liversedge, Stewart & Urbach, 1995). Thus, a child
who did not have an abstract representation of the passive structure could not be
primed to produce a passive sentence (e.g. The banana was eaten by the monkey). In
one priming study, Messenger et al. (2012) took turns describing pictures with
children as young as 3 years old. Contrary to previous findings that children up to the
age of 7 struggle to comprehend passive sentences with certain types of verb
(Maratsos, Fox, Becker & Chalkey, 1985), Messenger et al. (2012) found that
priming occurs in these young children for the passive structure, irrespective of the
verb type used in the prime sentence. Similar findings have been demonstrated with
18
passive, transitive and dative constructions (respectively, Bencini & Valian, 2008;
Huttenlocher, Vasilyeva & Shimpi, 2004; Thothathiri & Snedeker, 2008).
However, Savage, Lieven, Theakston and Tomasello (2003) demonstrated
that, while priming was present in the 3-year-olds they tested, this was only the case
when there was a large degree of lexical overlap between prime and target. That is,
when these young children were primed with a sentence containing pronouns (e.g. It
got pushed by it) they were able to produce a passive target (e.g. It got cut by it).
They were far less likely to produce a passive target when primed with a sentence
containing full noun phrases (e.g. The bricks got pushed by the digger), which could
not be reused in their own sentence. These results are therefore compatible with a
constructivist interpretation whereby the children are using a semi-abstract slot-and-
frame construction (e.g. It got Xed by it) to complete the priming task in the high
lexical overlap condition.
1.4 Summary
This chapter has discussed two opposing views of language acquisition: generativist-
nativist accounts and constructivist accounts. The literature reviewed here
demonstrate that both approaches appear to have some merit. On balance, however,
the generativist view is simply not able to explain critical aspects of language
development. Key amongst these are the apparent lexical effects that are readily
observable in children’s acquisition data. The ability of constructivist approaches to
explain these effects means this is the approach that must be adopted to explain the
by-verb differences in the pattern of results in the studies reported in this thesis.
Chapter 2 presents a review of the literature of particular interest:
overgeneralisation errors. Children observe patterns in the input, such as adding –ed
to the end of English verbs to create the past tense (walked, jumped, stopped, etc.).
This pattern can then usefully be generalised to other, newly acquired verbs to create
forms such as liked and dropped. However, children often over-apply these
generalisations and create ungrammatical, overgeneralised forms such as *breaked
and *sleeped. This type of morphological overgeneralisation error is frequently
found in children’s early language. The type of overgeneralisation under
investigation in this thesis is a little more complex: that of argument structure
overgeneralisation errors, such as *I filled toys into the box and *She giggled me.
19
1.5 The roles of authors in papers
Chapter 3 is a published paper (Bidgood et al., 2014). The conception and design of
the experiment were undertaken by all authors. Amy Bidgood collected the data,
performed initial analyses and wrote the first draft of the manuscript. Additional
analyses and interpretation of the data were undertaken in collaboration with the
other authors, as was revision of the manuscript.
Chapters 4 is a paper in preparation for submission (Bidgood, Ambridge,
Pine, Rowland & Freudenthal, in prep.). The conception and design of the
experiments were undertaken by the first four authors. The final author calculated the
frequency counts from the corpus. Amy Bidgood collected the experimental data,
performed initial analyses and wrote the first draft of the manuscript. Additional
analyses and interpretation of the data were undertaken in collaboration with
Ambridge, Pine and Rowland, as was revision of the manuscript.
Chapters 5 is a paper in preparation for submission (Bidgood, Ambridge, Pine
& Rowland, in prep.). The conception and design of the experiments were
undertaken by all authors. Amy Bidgood collected the experimental data, performed
initial analyses and wrote the first draft of the manuscript. Additional analyses and
interpretation of the data were undertaken in collaboration with the other authors, as
was revision of the manuscript.
21
Chapter 2: Argument structure overgeneralisation errors
2.1 Overgeneralisation errors
In English, certain pairs of constructions allow verbs to ‘alternate’ between them to
express related meanings. (Note that I am not suggesting that either construction is
derived from the other; see Goldberg, 2002). Good examples of alternating pairs are
the dative and locative constructions (1 and 2, respectively):
(1a) Paul gave Mary the book.
(1b) Paul gave the book to Mary.
(2a) Christine sprayed water onto the flowers.
(2b) Christine sprayed the flowers with water.
Becoming aware of these patterns is useful in language development. Initially,
children may observe that several verbs can alternate between a pair of constructions,
as with the following examples of the locative alternation:
(3a) Toby splashed the floor with water.
(3b) Toby splashed water onto the floor.
(4a) Ali loaded the car with bags.
(4b) Ali loaded bags into the car.
As children become productive language users, not restricted to simply imitating
utterances they have already heard, they are able to generalise the pattern to new
locative verbs (e.g. Ambridge, Pine & Rowland, 2012). After hearing a sentence such
as (5a), they may produce a novel sentence, such as (5b).
(5a) Frances packed the cupboard with food.
(5b) Frances packed food into the cupboard.
Becoming aware of patterns and generalising these patterns to new instances is a key
characteristic of human language use.
22
However, this process is not always straightforward. The problem of interest
in this thesis is how children know when to stop generalising a pattern. If they
assume that any verb heard in a ground-locative sentence, such as (5a), may also be
used in a figure-locative sentence, such as (5b), they are likely to produce erroneous
sentences such as those in (6):
(6a) *Howard poured the cup with tea.
(6b) *Becca spilt the carpet with juice.
(6c) *Glen spread the toast with peanut butter.
Errors such as these result from a process of generalisation and overgeneralisation,
and are therefore known as overgeneralisation errors.
Many diary studies have reported examples of argument structure
overgeneralisation errors of various different types (see Pinker, Lebeaux & Frost,
1987, for a summary of those reported in several previous papers). The following
examples are all taken from Bowerman (1996):
(7a) Dative: *I said her no. (c.f. examples 1a/b)
(7b) Locative: *Can I fill some salt into the bear [salt-shaker]? (c.f.
examples 2a/b)
(7c) Causative: *I saw a witch and she disappeared them.
Examples such as these, although not particularly common, attest to the fact that at
least some children go through a stage of producing argument structure
overgeneralisation errors, before correctly restricting their generalisations and
retreating from error. (N.B. Adults have also been reported to produce
overgeneralisations, although the source of these is unlikely to be the same as that of
children’s errors; see Pinker, 1989, pp. 154-160.)
The remainder of this chapter lays out the problem in more detail. Theoretical
approaches to children’s retreat from overgeneralisation errors, along with empirical
evidence, will then be presented, focusing on the three hypotheses that have most
informed the current thesis: the semantic verb class hypothesis (Pinker, 1989), the
entrenchment hypothesis (Braine & Brooks, 1995) and the preemption hypothesis
(Goldberg, 1995). As will become apparent, while evidence exists for all three
23
accounts, the data cannot be explained by any one theory alone. The final section of
the chapter therefore discusses research investigating semantic and statistical
approaches simultaneously, and accounts that integrate the various approaches.
2.2 Baker’s Paradox and the ‘no negative evidence’ problem
The difficulty in explaining children’s retreat from overgeneralisation errors was
termed Baker’s Paradox by Pinker (1989; see also Baker, 1979): children cannot
know that a certain verb cannot participate in an argument structure alternation
simply because they have not heard the verb used in that way. In addition, children
do not receive sufficient negative evidence to learn all of the exceptions to the rules
of English grammar. The following paragraph explains the ‘no negative evidence’
problem in more detail.
Brown and Hanlon’s (1970) classic study is the foundation upon which the
‘no negative evidence’ claim is based. They found that the parents of the children
they studied were no less likely to accept their child’s utterance (in terms of how they
responded to it) if it was ungrammatical than if it was grammatical. The focus of
parents’ responses was on the semantics, not the syntax. However, several follow-up
studies have questioned this original finding. Hirsh-Pasek, Treiman and
Schneiderman (1984) replicated Brown and Hanlon’s finding that parental
acceptance did not vary with grammaticality of child utterance but carried out an
additional analysis, illustrating that parents of 2-year-olds (although not older
children) were more likely to repeat or recast their child’s utterance if it was
ungrammatical than if it was grammatical (see also Chouinard & Clark, 2003;
Demetras, Nolan Post & Snow, 1986). Penner (1987) also found that parents were
more likely to recast or expand on children’s ungrammatical utterances (and, again,
found that this was more frequent in parents of younger children), whereas they were
more likely to move the topic on following a grammatical utterance. These findings
show that parents are certainly sensitive to the grammaticality of their children’s
utterances.
In contrast to Brown and Hanlon’s (1970) study, then, follow-up research has
shown that some children do receive feedback on grammatical errors, albeit often in
the form of indirect negative evidence. However, this cannot be the case for all
utterances produced by all children (in the case of utterances produced only in the
24
company of other young children, for example). Perhaps more important is the
question of whether children are able take on board the feedback they receive. Below
is one of several conversations reported by Pinker (1989, p. 13; from Braine, 1971)
illustrating that children are, at least in some cases, unable to take on board either
implicit (lines 2 and 3) or explicit feedback:
(8) Child: Want other one spoon, Daddy.
Father: You mean, you want THE OTHER SPOON.
Child: Yes, I want other one spoon, please, Daddy.
Father: Can you say “the other spoon”?
Child: Other… one… spoon.
Father: Say… “other.”
Child: Other
Father: “Spoon.”
Child: Spoon.
Father: “Other… spoon.”
Child: Other… spoon. Now give me other one spoon?
Examples such as (8) suggest that parental feedback alone cannot account for
children’s retreat from overgeneralisation.
However, Saxton and colleagues have found that some forms of feedback do
seem to reduce children’s production of ungrammatical utterances. Saxton, Kulcsar,
Marshall and Rupra (1998) conducted an experimental study, testing children at two
time points, five weeks apart. They found that children who had received negative
evidence on the over-regularisation of irregular past tense verbs at the first time point
(as opposed to positive input) were more accurate at the second time point. Saxton,
Backley and Gallaway (2005) also found improvement over a 12-week period, but
only for three of the 13 types of grammatical error they examined. In contrast to the
positive effect suggested by Saxton’s work, for at least some error types, Morgan,
Bonamo and Travis (1995) found no evidence that recasts led to children producing
more grammatical utterances and, in the long term, the number of recasts was
actually a negative predictor of grammaticality (see also Morgan & Travis, 1989).
The above studies provide mixed support for the role of negative evidence in
children’s retreat from the production of various error types. It therefore follows that
25
negative evidence is unlikely to be the primary means via which children retreat from
overgeneralisation errors, and therefore this unlikely to be a solution to Baker’s
Paradox. The remainder of Pinker’s (1989) book (see also an earlier account in
Pinker, 1984) is devoted to solving Baker’s Paradox, laying out an account
suggesting that children are able to use semantic information to retreat from, or
avoid, argument structure overgeneralisation errors. This semantic verb class
hypothesis is discussed in depth in section 2.3.1.
2.3 Theoretical approaches to explaining the retreat from overgeneralisation
This section discusses the three theoretical approaches which have been most
influential to this thesis, and in the field more generally: the semantic verb class
hypothesis (Pinker, 1989), the entrenchment hypothesis (Braine & Brooks, 1995) and
the preemption hypothesis (Goldberg, 1995). Each subsection will give an overview
of the theory itself as well as presenting empirical evidence in support of that theory.
Section 2.4 discusses recent attempts to integrate these three theories into an
approach that better explains all of the data.
2.3.1 The semantic verb class hypothesis
Pinker’s (1989) semantic verb class hypothesis involves a mechanism whereby fairly
broad semantic structures, known as thematic cores, are linked to particular verb
argument structures via innate linking rules. Numerous verbs may have the same
thematic core, and each verb has a separate lexical entry for each thematic core that
relates to it. Using the locative constructions as an example, the thematic cores for
the figure locative (e.g. Christine sprayed water onto the flowers) and ground
locative (e.g. Christine sprayed the flowers with water), respectively, are as follows
(p. 77):
(9a) X moves Y into/onto Z
(9b) X causes Y to change its state by means of moving Z to Y
Thus, verbs such as spray, which are able to alternate between the two locative
constructions, have one lexical entry denoting that a substance is moved to a location
26
via the method of spraying and second lexical entry denoting that moving a
substance to a container or location (the flowers, in the example above), via a
spraying motion, has changed the state of that container/location (the flowers are
completely sprayed with water). Relating the two locative constructions allows
children to set up a broad-range rule linking them together (although the set of
possible rules is constrained by the innate linking rules that Pinker proposes).
Some verbs have only one lexical entry for the locative construction. Pour is
a figure-only verb (cf. Howard poured tea into the cup/*poured the cup with tea), so
has a lexical entry related to (9a) but not (9b). In contrast, fill is a ground-only verb
(cf. Howard filled the cup with tea/*filled tea into the cup) and this has a lexical
entry related to (9b) but not (9a). If children posit that all verbs that can appear in one
of these two constructions are able to appear in the other, as is the case for spray but
not pour or fill, overgeneralisation errors will result.
To solve the problem of why only some verbs are able to alternate between
two related constructions, like the ground and figure locative, Pinker proposed that
there are, in fact, more specific, narrow semantic classes (sometimes referred to as
subclasses) within each of the broad semantic classes. Some of the narrow classes
allow alternation between the two locative constructions, via a narrow-range rule,
whereas others do not. According to Pinker (1989, pp. 126-7) spray belongs to the
alternating narrow class defined as “force is imparted to a mass, causing ballistic
motion in a specified spatial distribution along a trajectory”. As children learn other
verbs with semantics fitting this definition, such as splash, sprinkle and squirt, they
know that both locative constructions are licensed. Pour belongs to the figure-only
narrow class defined as “a mass is enabled to move via the force of gravity”. As
children learn other verbs with semantics fitting this definition, such as drip, shake
and spill, they know that only the figure locative is licensed. Fill belongs to the
ground-only narrow class defined as “a layer completely covers a surface”. As
children learn other verbs with semantics fitting this definition, such as cover, line
and pad, they know that only the ground locative is licensed.
In contrast to the broad semantic classes and broad range rules, then, once the
semantics of a particular verb have been learnt sufficiently well to place it into one of
the narrow semantic classes, a child will know whether or not the alternative
construction is licensed. This allows for productivity in language use, since a child
need only hear an alternating verb in one of the constructions (or, in fact, in neither
27
construction; see Chapter 3) to be confident in using it in the other. In order to retreat
from error, the narrow semantic classes and narrow-range rules must be properly
learnt; until this is the case, children’s language will be error-prone.
The studies of Gropen, Pinker, Hollander and Goldberg (1991a, b) used the
locative construction to test the predictions of Pinker’s semantic verb class
hypothesis. Gropen et al. (1991a) conducted experiments using both elicitation and
comprehension methodologies to test the prediction of the semantic verb class
hypothesis that lack of detailed knowledge of verb semantics is the root cause of
overgeneralisation errors. According to the innate linking rule of object affectedness,
the direct object of the verb is the one which is affected by an event. Thus, in the
figure-locative sentence Howard poured tea into the cup, the tea is most affected,
having been moved from its original location in the teapot to its new location in the
cup. In contrast, in the ground-locative sentence Mark filled the cup with tea, the cup
is most affected as it has changed from being empty to being full.
To test participants’ knowledge of verb semantics, Gropen et al. (1991a) used
a forced-choice pointing task with a series of line drawings. Participants were first
introduced to the illustrations for each of the verbs in question using an ambiguous
sequence of pictures. For example, a woman pouring water from a jug to a glass,
which ends up full, could be accurately described as a pouring or a filling event.
They were told that this was either pouring or filling (they received trials with each
verb). The ambiguous picture was then replaced with two unambiguous ones: the
pouring picture showed a woman pouring water but the water spilling, so the glass
remained empty; the filling picture showed a woman dripping water from a tap into
the glass, which ended up full. Participants were asked to point to pouring or filling,
whichever was the target verb on that trial. While adults performed significantly
above chance for all verbs, children (aged 2-5 years) were only able to do so
consistently with figure-only pour-type verbs, often misinterpreting fill as meaning
something more like pour.
To test participants’ knowledge of verb argument structures, participants
were asked what the woman was doing to the water (figure/contents-focussed
question) or the glass (ground/container-focussed question) in order to elicit verbal
descriptions using each target verb. Again, adults almost always produced
grammatical utterances to describe the events. However, children were more likely to
produce ungrammatical forms, particularly using fill in the figure locative
28
construction (e.g. *She’s filling it into the glass in response to the question, What’s
the woman doing to the water?). Older children (aged 4;6-5;11) also produced fewer
ungrammatical responses than younger children (2;6-4;5). Gropen et al. (1991a) did
not find evidence of contingencies between semantic knowledge and error rates in
their first experiment (just described), but repeated the study focussing solely on the
verb fill, since this seemed to cause the most problems in terms of interpretation.
This time, for children aged 3;6-6;6, results showed that children who were biased to
interpreting the meaning of fill as being related to a manner (e.g. pouring) as opposed
to an end-state (being full) were significantly more likely to produce ungrammatical,
figure-locative utterances containing fill, than those who were not biased to this
manner interpretation. Thus, these results support the semantic verb class hypothesis,
albeit on a very limited scale: better semantic knowledge about the verb fill led to
fewer overgeneralisation errors involving that verb.
In a further test of the semantic verb class hypothesis, Gropen et al. (1991b)
tested participants’ ability to categorise new verbs in line with Pinker’s (1989) broad
semantic classes and whether children (aged 3-9 years) and adults would be able to
use this knowledge to produce sentences containing the novel verbs in the expected
locative construction. Participants were taught manner-of-motion verbs (e.g. keating
= moving in a zig-zag motion), designed to be interpreted as a figure-only verb, like
pour. If participants were able to interpret these verbs as intended, they should
produce more figure-locative than ground-locative responses (e.g. You’re keating the
marble to the cloth, rather than You’re keating the cloth with the marble).
Participants were also taught end-state verbs (e.g. mooping = changing colour),
designed to be interpreted as a ground-only verb, like fill. If participants were able to
interpret these verbs as intended, they should produce more ground-locative than
figure-locative responses (e.g. You’re mooping table with the cloth, rather than
You’re mooping the cloth onto the table).
Results showed that both adults and children were able to use the new verbs
as expected, based on manner-of-motion or end-state interpretations. This provides
strong evidence for the semantic verb class hypothesis: children and adults were able
to use verb semantics alone (participants only heard the verb in its gerund form in the
teaching phase) to identify the correct locative construction with which to produce a
sentence (based on other verbs with broadly similar semantics). Using semantic
29
knowledge to restrict argument structure generalisations when learning real verbs is
therefore highly plausible.
2.3.2 The entrenchment hypothesis
The entrenchment hypothesis was first proposed by Braine and Brooks (1995) as a
mechanism by which children may avoid or retreat from overgeneralisation errors.
Entrenchment is an inference-from-absence mechanism. As children hear more and
more instances of a particular verb being used, they are able to track the statistical
distribution of that verb in the input. The more a particular verb is heard in a
particular construction, the more that verb becomes entrenched in that construction.
Of course, some verbs may be heard in several different constructions:
(10a) Ruby cut the cake.
(10b) Ruby cut the cake with a knife.
(10c) The cake was cut.
(10d) The cake was cut by Ruby.
(10e) This cake cuts easily.
(10f) Ruby cut me a slice of cake.
(10g) Ruby cut a slice of cake for me.
(10h) Cut the cake!
Children hear the verb cut used in different constructions, such as those in (10), but
they are highly unlikely to hear the verb use in an ungrammatical sentence such as
*The cake cut into pieces. Having heard cut in numerous constructions, but never in
this ungrammatical intransitive construction, children might infer that it is not
possible for cut to be used this way. Realising that this is the case enables children to
retreat from, or avoid, overgeneralisation errors, such as *The cake cut into pieces.
Since different verbs are heard with different frequencies in the input, the
entrenchment hypothesis predicts by-verb differences in error rates. According to this
hypothesis, children are less likely to produce errors with a verb the higher its
frequency in the input, as the inference-from-absence is strengthened each time a
verb is heard. Similarly, the more frequently a verb has been heard in the input, the
30
less likely that verb is to be judged as being grammatically acceptable when it is
heard in an overgeneralised construction.
Brooks, Tomasello, Dodson and Lewis (1999) examined the predictions of
the entrenchment hypothesis using the causative alternation. Some verbs are able to
alternate between the transitive-causative construction (e.g. Robert rolled the ball)
and the intransitive-inchoative construction (in which no external agent is expressed,
e.g. The ball rolled). Others are only grammatical in the transitive-causative (cf.
Robert hit the ball/*The ball hit) while others still are only grammatical in the
intransitive-inchoative (cf. The girl laughed/*Robert laughed the girl). Brooks et al.
tested the prediction of the entrenchment hypothesis that children will make fewer
errors with verbs that are more frequent in the input (i.e. those that are more
entrenched in their grammatical constructions and for which a stronger inference-
from-absence is therefore available). Children observed puppets performing 4
different actions. These actions were each described with two different verbs: one
with early age of acquisition (AOA) and one with late AOA. (Verbs with an early
AOA are likely to be more frequent in the input than those with a late AOA.) Two
verb pairs were transitive-only (e.g. hit and strike; early and late AOA, respectively)
and two were intransitive-only (e.g. come and arrive; early and late AOA,
respectively).
Participants were asked a series of questions to elicit descriptions of the
events containing the target verbs. Questions were either agent-focussed (e.g. What is
the [agent] doing?; designed to encourage production of transitive sentences),
patient-focussed (e.g. What is happening with the [patient]?; designed to encourage
production of intransitive sentences) or neutral (e.g. What is happening now?). The
results were in line with the predictions of the entrenchment hypothesis: children
produced more than twice as many overgeneralised responses with late AOA than
early AOA verbs. This study therefore provides strong support for the entrenchment
hypothesis. It is also interesting to note that children produced, on average, three
times as many erroneous transitive sentences with intransitive-only verbs as they did
erroneous intransitive sentences with transitive-only verbs.
Theakston (2004) employed a grammaticality judgment methodology to test
the extent to which adults and children (aged 5 and 8 years) would accept
overgeneralisation errors with high- and low-frequency verbs as being grammatically
acceptable. Children completed a forced choice task (acceptable or unacceptable),
31
whereas adults provided their judgments on a 7-point scale (1 = completely
acceptable, 7 = completely unacceptable). Participants of all ages heard sentences
from various different construction pairs, some of which used verbs in their generally
accepted argument structure, whilst the other contained the same verb in an
overgeneralised structure (e.g. It fell off vs. *Somebody fell it off). For each high-
frequency verb (e.g. fall), a low-frequency equivalent was chosen (e.g. tumble), and
similar sentence pairs were constructed.
Results showed that children of both age groups were significantly more
likely to accept overgeneralisation errors with low- than with high-frequency verbs.
In addition, the 5-year-olds were more likely than the 8-year-olds (who had
presumably had more exposure to the verbs in question in grammatical
constructions) to accept overgeneralisation errors. Finally, the results for the adults
were in line with those of the children, with adults judging errors with low-frequency
verbs as being more acceptable than those with their high-frequency equivalents.
Overall, then, the results from Theakston’s (2004) study again provide strong support
for the entrenchment hypothesis, this time employing a judgment methodology and
investigating a wide range of constructions.
2.3.3 The preemption hypothesis
The preemption hypothesis (e.g. Goldberg, 1995) incorporates elements of both
statistical and semantic information. In common with entrenchment, preemption
involves hearing instances of a verb in its authorised constructions and using this
information to infer that certain other constructions are not compatible with that verb.
The critical difference is that one construction can only preempt the use of an
alternative, ungrammatical construction if the two are roughly equivalent in meaning.
In the following example, sentences such as (11a), but not (11b), preempt (and
therefore help children avoid) the error in (11c):
(11a) Howard poured tea into the cup.
(11b) Water poured out of the burst pipe.
(11c) *Deborah poured the pan with oil.
32
In a similar way to the entrenchment hypothesis, hearing a verb like pour used many
times in the figure-locative construction (such as 11a), a construction with similar
semantics to the ungrammatical ground-locative construction (such as 11c), allows
children to infer that it cannot be used in the latter construction because the relevant
meaning is conventionally expressed using the former. Having a readily available,
alternative construction that is semantically very similar thus allows children to avoid
errors such as (11c).
Like the entrenchment hypothesis, by-verb differences are also predicted by
the preemption hypothesis. In this case, the more frequently a verb has been heard in
the grammatical construction that preempts the erroneous one, the less likely children
are to produce that verb in the ungrammatical construction.
To test the predictions of the preemption hypothesis in a controlled manner,
Brooks and Zizak (2002) taught children (aged 4 and 6-7 years) two novel verbs. The
verbs each described a novel action similar in semantics to Pinker’s (1989) manner-
of-motion verbs (e.g. roll). Importantly for the purposes of this study, these verbs are
able to alternate between the transitive and intransitive constructions (e.g. The ball
rolled/Robert rolled the ball) and, thus, the novel verbs could be used
‘grammatically’ in either construction (e.g. The tree is tamming/The mouse is
tamming the tree). Participants were assigned to one of three conditions: the
Alternative Construction group; the English Suppletive group; or the No Preemption
group. All participants heard one action described with a novel verb in the
intransitive construction (e.g. The tree is tamming) and another with a novel verb in
the transitive construction (e.g. The rabbit is dacking the car). Participants in the
Alternative Construction group also heard the intransitive novel verb in the
preempting periphrastic causative construction (e.g. The mouse is making the tree
tam) and the transitive novel verb in the preempting passive construction (e.g. The
car is getting dacked). In contrast, as well as hearing the novel verbs in the
intransitive or transitive constructions, participants in the English Suppletive group
heard the actions described with real (English) verbs (e.g. swing, bounce) with the
opposite transitivity. Thus, participants in both the Alternative Construction group
and the English Suppletive group were able to provide the agent of the intransitive
action and omit the agent of the transitive action whilst keeping the transitivity of the
novel verbs as those verbs had been taught. In contrast, participants in the No
Preemption group heard no alternatives and thus, if the preemption hypothesis is
33
correct, would be more likely to alter the transitivity of the novel verbs in order to
provide or omit the agent in the intransitive and transitive conditions, respectively.
As in Brooks et al. (1999), participants were encouraged to use the novel
verbs through a series of agent-focussed, patient-focussed and neutral questions.
Participants’ responses were coded for transitivity. The 4-year-old children in the
English Suppletive group were significantly less likely to violate the assigned
transitivity of the novel verbs than those in the No Preemption group (there was no
difference between those in the Alternative Cosntruction group and the No
Preemption group for these younger children). The 6- to 7-year-old children in both
the English Suppletive and Alternative Construction groups were significantly less
likely to violate the assigned transitivity of the novel verbs than those in the No
Preemption group. These results therefore show support for the preemption
hypothesis, particularly for the older children, as those given an alternative means of
providing the agent for an intransitive novel verb or omitting the agent for a
transitive novel verb were less likely to generalise those verbs into the alternative
construction, even though the verb’s semantics might lead the participants to believe
that this was possible.
Goldberg (2011) used corpus evidence to test whether the probability of
hearing a verb in one construction, when discourse context might have lead the
language learner to think a second construction would have been more appropriate, is
sufficient to infer that that second construction is, in fact, ungrammatical. By
expressing the preemption hypothesis in this way, Goldberg was able to test the
evidence for an explicit mechanism of how preemption might operate, using
evidence from the input (or, at least, using corpus evidence as a proxy for this). The
alternation examined in this paper was the dative:
(12a) Double object [DO] dative: Mel told Sue the news.
(12b) Prepositional object [PO] dative: Mel told the news to Sue.
Not all verbs are grammatical in the DO dative: *Mel shouted Sue the news (cf. Mel
shouted the news to Sue). These are PO-only verbs. (Note that some verbs are
grammatical in the DO but not the PO dative [Mel bet Sue £5 vs. *Mel bet £5 to Sue],
but these were not examined by Goldberg.) To test her theory, Goldberg took the
situation in which the recipient of an action (Sue in 12) was expressed by a pronoun,
34
but the theme (the news in 12) was expressed with a full NP. In this specific context,
the DO dative is generally preferred (i.e. Mel told her the news > Mel told the news
to her). This context therefore provides a good test case for Goldberg’s theory: if a
dative verb with a pronominal recipient and full NP theme is heard in the PO dative
more than in the DO dative, it suggests to learners that the DO is not grammatical.
Goldberg’s corpus evidence showed that, using this statistic (the probability
of hearing a dative verb with a pronominal recipient and full NP theme in the PO
dative, e.g. Mel told the news to her), the input alone provided enough evidence for
language learners to be able to discriminate statistically between alternating and PO-
only verbs. For alternating verbs, such as tell, the probability of hearing a verb with a
pronominal recipient and full NP theme in the PO dative was, on average, 0.04. In
comparison, for PO-only verbs, such as explain, the average probability was 0.83.
For example, all else being equal, children might expect to hear explain in the DO
dative (with pronominal recipient and full NP theme) around 96% of the time, like
other dative verbs. However, they actually hear explain in this context in only around
1% of cases (Goldberg, 2011, p. 137). They can therefore conclude that the PO use
must be ungrammatical: the DO uses in this unexpected context preempt the
ungrammatical PO use of the verb explain. Thus, on the basis of discourse context
and distributional information, this test case shows that children would be able to use
the preemption mechanism to learn which dative verbs are able to alternate between
the PO and DO dative and which are only grammatical in the PO dative.
2.4 Integrating semantic and statistical approaches
As demonstrated in section 2.3, the semantic verb class hypothesis (Pinker, 1989),
the entrenchment hypothesis (Braine & Brooks, 1995) and the preemption hypothesis
(Goldberg, 1995) all enjoy a certain amount of empirical support. However, none of
these theories is capable of explaining all of the data: the semantic verb class
hypothesis cannot explain by-verb differences related to input frequency, and neither
the entrenchment hypothesis nor the preemption hypothesis is able to explain how
children are apparently able to use verb semantics to select the correct construction in
which to use a novel verb without having ever heard it in a sentence. This section
first presents several studies that have simultaneously examined the predictions of
two or more of the semantic verb class, entrenchment and preemption hypotheses. It
35
then describes one account, the FIT account, that attempts to integrate aspects of all
three theories in order to provide a more complete picture of the mechanism(s)
involved in children’s retreat from overgeneralisation.
Brooks and Tomasello (1999) used a novel-verb elicitation task, similar to
Brooks et al. (1999) and Brooks and Zizak (2002), to test the predictions of both the
preemption and semantic verb class hypotheses with the causative alternation (The
ball rolled/Robert rolled the ball). Children were taught two novel verbs. For half of
the children, meek was taught as a transitive-only verb of causation of directed
motion (e.g. raise). These children heard meek used only in transitive contexts (e.g.
The mouse is meeking the flower). They were also taught tam, an alternating manner-
of-motion verb. Children in this group heard tam only in intransitive contexts (e.g.
The car is tamming). For the other half of the children, meek was taught as an
intransitive-only verb of inherently directed motion (e.g. rise). These children heard
meek used only in intransitive contexts (e.g. The flower is meeking). Again, they
were also taught tam, an alternating manner-of-motion verb. However, children in
this group heard tam only in transitive contexts (e.g. The doll is tamming the car).
This enabled Brooks and Tomasello to test the semantic verb class hypothesis: meek
always belonged to a non-alternating class, and the semantic verb class hypothesis
predicts children’s production of this verb would be limited to the construction in
which it was modelled; tam always belonged to the alternating manner-of-motion
verb class, and the semantic verb class hypothesis predicts children would produce
this verb in both constructions, in appropriate contexts.
To test the preemption hypothesis, in a similar way to Brooks and Zizak
(2002), half of the children (in the No Preemption group) heard each verb in only one
context: transitive or intransitive. The other half of the children (in the Preemption
group) heard each verb in two contexts: transitive and truncated passive (e.g. The
mouse is meeking the flower and The flower is getting meeked) or intransitive and
periphrastic causative (e.g. The car is tamming and The doll is helping the car tam).
This gave children in the Preemption group a way to place the discourse focus of
either verb on the agent or patient of the action without changing the verb’s
transitivity. Children in the No Preemption group were, of course, able to use
alternative structures in their own production if they wished, but they had not heard
the novel verbs modelled in this way.
36
Throughout the test sessions, children were asked agent-focussed, patient-
focussed and neutral questions to elicit a range of responses from the children.
Results showed that, overall, children tended to produce verbs with the transitivity in
which they had been heard in training. In support of the semantic verb class
hypothesis, children aged 4 and 6-7 years (although not children aged 2 years) were
significantly more likely to produce sentences with the opposite transitivity for tam
(from the alternating manner-of-motion verb class) than for meek (from the fixed-
transitivity verb classes). In support of the preemption hypothesis, the oldest children
(aged 6-7 years) in the No Preemption group produced significantly more responses
of the opposite transitivity to that heard in training than did the children of the same
age in the Preemption group. This effect was not found for the younger children.
Brooks and Tomasello therefore conclude that both semantic verb class and
preemption have an effect on children’s overgeneralisation errors, and that this effect
increases with age.
Again using the causative alternation (The ball rolled/Robert rolled the ball),
Ambridge et al. (Ambridge et al., 2008; Ambridge, Pine, Rowland, Jones & Clark,
2009; Ambridge, Pine & Rowland, 2011) ran a series of experiments using a
grammaticality judgement methodology to test semantic and statistical accounts
simultaneously. In grammaticality judgment studies, participants typically watch an
animation depicting an event. They then hear a sentence describing the event, which
may, or may not, be generally considered to be grammatical. For example, they may
see an event in which a boy tickles a girl and the girl laughs and then hear a
description such as The girl laughed or *The boy laughed the girl. Participants
provide a grammaticality judgement rating for the sentence. Ambridge et al. typically
asked children to provide their judgements on a 5-point ‘smiley-face’ scale (see
Figure 2.1), whereas adults provided judgements either using the same scale or a
simple numerical scale (1-5 or 1-7, where 1 is completely ungrammatical and 5 or 7
is completely grammatical). In some studies, adults did not watch animations but
judged written sentences instead.
37
Figure 2.1. 5-point ‘smiley-face’ scale used in Ambridge et al.’s grammaticality
judgement studies, where the face on left represents a completely ungrammatical
sentence, the face on the right a completely grammatical sentence, and the remaining
faces a rating between these two extremes.
Ambridge et al. (2008) collected grammaticality ratings from children (aged
5-6 and 9-10) and adults for grammatical intransitive sentences (e.g. Bart fell into a
hole) and ungrammatical transitive sentences (e.g. *The man fell Bart into a hole)
containing either real or novel verbs. To test Pinker’s semantic verb class hypothesis,
the verbs were chosen from classes with a low degree of external causation (e.g.
verbs of semi-voluntary expression of emotion; laugh) and with a higher degree of
external causation (the prototypical meaning of the transitive-causative construction;
e.g. verbs of going out of existence; disappear). In support of this hypothesis,
participants judged transitive overgeneralisation errors with a novel verb from the
class with a low degree of external causation (e.g. *The man laughed the boy) to be
less grammatical than such errors with a novel verb from a class with a higher degree
of external causation (e.g. *The man disappeared the rabbit). To test the
entrenchment hypothesis, verbs of high and low frequency in the input, along with
novel verbs, were included. In support of this hypothesis, participants judged
transitive overgeneralisation errors with high-frequency verbs (e.g. *The man
laughed the boy) to be significantly worse, in comparison with their grammatical
intransitive equivalents (e.g. The boy laughed), than such errors with low-frequency
verbs (e.g. *The man giggled the boy), in comparison with their grammatical
intransitive counterparts (e.g. The boy giggled). Ambridge et al. (2008) concluded
that both the formation of semantic verb classes and an entrenchment mechanism
play a role in children’s retreat from overgeneralisation errors.
Ambridge et al. (2011) used a similar method, although, this time, verbs were
either intransitive-only (as in Ambridge et al. 2008) or alternating (e.g. The ball
38
rolled/The man rolled the ball). This had the advantage that participants could not
use a strategy of consistently judging transitive sentences to be ungrammatical to
give their responses. Children in this study judged sentences containing only novel
verbs, whilst adults judged sentences containing both novel and real verbs.
Consistent with earlier findings supporting the semantic verb class hypothesis,
participants of all ages judged sentences containing novel verbs to be grammatical or
ungrammatical in line with equivalent sentences containing known verbs from the
same semantic class. Consistent with earlier findings supporting the entrenchment
hypothesis, adults judged high-frequency real verbs to be significantly less
grammatical in the ungrammatical transitive sentence than equivalent sentences
containing their low-frequency counterparts. Ambridge et al. (2011) also
manipulated the frequency of the novel verbs taught to adults, but no effect of
entrenchment was observed with these verbs. The conclusion, again, was that both
the formation of semantic verb classes and an entrenchment mechanism play a role in
children’s retreat from overgeneralisation errors.
Ambridge et al. (2011; see also Ambridge & Lieven, 2011) proposed the FIT
account as an attempt to integrate semantic and statistical accounts. (Note that other
accounts that integrate semantic and statistical elements have also been proposed,
including Langacker [2000], MacWhinney [2004] and Tomasello [2003].) Under the
FIT account, the sentence a speaker produces depends on various factors. These
factors include the frequency of both the construction itself (higher-frequency
constructions are more highly activated) and a particular verb in that construction,
and the relevance of the construction in conveying the message. Frequency accounts,
both entrenchment and preemption, are therefore accommodated under this account.
The key to the FIT account, though, and the factor that gives it its name, is
the semantic compatibility (or ‘fit’) between the lexical items and the empty slots in
the construction. For example, the slots in the transitive-causative construction are
[AGENT] [ACTION] [PATIENT]. The prototypical semantics of these slots are
gradually acquired through exposure to the input. Over time, children learn that the
verbs that appear in the [ACTION] slot in transitive-causative construction are
related to direct, external physical causation (e.g. kick). That means that new verbs
children learn that also relate to direct, external physical causation are also
compatible with the [ACTION] slot in the transitive-causative construction, and
children will be able to deduce that they can use these verbs in this construction. In
39
contrast to the transitive-causative construction, the intransitive-inchoative
construction contains the slots [PATIENT/ACTOR] [ACTION], where the
prototypical semantics of the [ACTION] slot relate to internal causation. Again, the
semantics of this slot are learnt over time, through exposure to the input (e.g. The
man laughed, The fish swam, The hot air balloon rose). When children learn new
verbs that also denote internal causation, they will be able to slot them into the
[ACTION] slot in the intransitive-inchoative construction.
Some verbs, such as roll and open, have properties that relate to both internal
and external causation. For example, it is only possible to roll a ball because of its
round shape, and it is only possible to open a box because its lid allows us to do so.
(Only unusual boxes would be able to be rolled, and only very unusual balls would
be able to be opened.) Nevertheless, these actions usually require some external force
to enable the actions to take place. Thus, verbs such as roll and open are semantically
compatible with both the transitive-causative and intransitive-inchoative
constructions and can be used in both.
The same argument applies to other constructions. For example, in the
locative alternation (ground locative = Frances filled the cupboard with food; figure
locative = Frances placed food into the cupboard), the ground locative construction
has the following construction slots: [AGENT] [ACTION]
[CONTAINER/LOCATION] with [CONTENTS]. In this construction, the semantics
of [ACTION] slot relate to a change of state of the [CONTAINER/LOCATION]
(e.g. the cupboard ends up completely full of food in the sentence Frances filled the
cupboard with food); it is irrelevant how Frances moved the food to its new location.
In contrast, the construction slots for the figure locative are [AGENT] [ACTION]
[CONTENTS] into/onto [CONTAINER/LOCATION]. The semantics of the
[ACTION] slot in this construction relate to the manner of motion of the contents
when it is being transferred into the container. So, the manner in which the food
enters the cupboard is different if it is placed/thrown/shoved into it; the cupboard
may or may not end up in a full state.
As with the transitive-causative and intransitive-inchoative constructions,
some verbs (e.g. pack, spray) have semantics that are compatible, to a certain extent,
with the [ACTION] slots in both constructions and are, thus, able to be used
grammatically in both. Nevertheless, the choice of construction is not irrelevant to
the meaning of the resulting sentence – the ground and figure locative constructions
40
themselves carry meaning. For example, the ground locative sentence Christine
sprayed the wall with paint implies that the entire wall ended up covered with paint,
whereas the figure locative sentence Christine sprayed paint onto the wall does not.
Thus, just because a verb is compatible with the [ACTION] slot in both constructions
does not mean that both constructions are equally good choices to convey the
message: if a change of state of the [CONTAINER/LOCATION] is important to the
message, the ground locative is likely to be the best choice.
Importantly, the production of overgeneralisation errors can also be explained
by the FIT account. If a child wants to express the causer of a laughing event, the
grammatical choice would be the periphrastic causative (The boy made the girl
laugh). However, this is a low-frequency construction and therefore has only low
levels of activation. (Note that ‘level of activation’ here refers to the baseline level of
accessibility of the construction in memory.) The intransitive-inchoative construction
is high frequency (The girl laughed), but does not convey all aspects of the message.
While the verb laugh is not semantically compatible with [ACTION] slot in the
transitive-causative construction (no external causation is required by the verb
laugh), the frequency of the construction itself and its ability to convey the entire
message mean that this disadvantage may be overlooked, and the ungrammatical
sentence *The boy laughed the girl may be produced.
Ambridge and colleagues (Ambridge, Pine & Rowland, 2012; Ambridge,
Pine, Rowland & Chang, 2012; Ambridge et al., 2014) have since used the
grammaticality judgement methodology to test the predictions of semantic and
statistical accounts, and the FIT account, with other constructions: the locative and
the dative. By using mixed effect models to analyse the data, they were able to
investigate the relative importance of the various semantic and statistical factors of
interest. In addition, this allowed for the use of actual verb frequencies (both overall
and in the relevant constructions) based on corpus counts, rather than relying on
general high versus low frequency groups of verbs.
Ambridge, Pine and Rowland (2012) investigated the locative alternation
(figure locative: Lisa sprayed water onto the flowers; ground locative: Lisa sprayed
the flowers with water). While some locative verbs (e.g. spray) can appear in both
the figure- and ground-locative constructions, others are figure-only (e.g. pour: Lisa
poured water into the cup/*Lisa poured the cup with water) or ground-only (e.g. fill:
*Bart filled water into the cup/Bart filled the cup with water). (Note that this study
41
used only known verbs and did not include novel verbs, which would have provided
a stronger test of the semantics hypothesis.) Findings suggested that the data were
best explained by a model that included both broad- and narrow-range semantic rules
(see section 2.3.1, above) and overall verb frequency (entrenchment). Verb
frequency in the preempting construction (in this case, the alternative locative
construction) was found to have no dissociable effect from overall verb frequency,
and therefore no support was found for a separate preemption mechanism. Although
the paper reports support for Pinker’s (1989) semantic classes, the semantic ratings
collected in this study actually suggest that semantics should not be seen as
categorical at all but, rather, should be viewed as a continuum. Variability in ratings
also support this notion: sentences were not judged in a binary way as being
grammatical or not, but on a scale, with some being slightly more or less
grammatical than others. Thus, the probabilistic nature of the FIT account, outlined
above, seems to capture the data better than Pinker’s class-based proposal.
Ambridge et al. (2014) tested the predictions of the semantics, entrenchment
and preemption hypotheses using the dative alternation (e.g. Bill gave a present to
Sue/Bill gave Sue a present, c.f. I said no to her/*I said her no). This study built on
the new approach to semantics used in Ambridge, Pine and Rowland (2012). Instead
of relying on the predetermined semantic classes proposed by Pinker (1989),
Ambridge, Pine and Rowland collected semantic ratings from a group of adults (who
did not participate in the grammaticality judgement task). Adults rated verbs based
on the semantic features suggested by Pinker as being important for this alternation,
but the researchers used Principal Components Analysis to produce a set of objective
semantic criteria. This method provided a graded semantic measure, rather than the
discrete classes of Pinker’s original proposal. Findings showed that graded verb
semantics predicted participants’ grammaticality judgments. Thus, although Pinker’s
notions of the important semantic features for this alternation are likely to be correct
(as the features against which verbs in this study were rated are those suggested by
Pinker), the notion of discrete semantic classes is not the best way to capture these
features. In terms of frequency measures, both entrenchment and preemption had
dissociable effects on participants’ grammaticality judgements. Unlike the case of the
locative alternation, then, preemption does seem to play a role, over and above that
of entrenchment, in the retreat from overgeneralisation errors with the dative
construction. This finding is important in itself, as it indicates that different
42
mechanisms may be used to different extents in the retreat from error for different
constructions. Overall, the model that best fits the data once again includes semantic
and frequency information. Thus, an integrated account, such as the FIT account, is
likely to be the way forward in terms of thinking about how children retreat from
overgeneralisation errors.
Drawing together the findings of the grammaticality judgement studies of
Ambridge and colleagues, a successful account must include roles for both semantic
and statistical information, but must also be flexible enough to account for
differences between different constructions. The FIT account fulfils all of these
criteria.
2.4.1 Distinguishing between entrenchment and preemption
The discrepancies in the findings of Ambridge, Pine and Rowland (2012) and
Ambridge et al. (2014) in terms of the roles of the entrenchment and preemption
mechanisms deserves further comment. Entrenchment is an inference-from-absence
mechanism: when a verb has been heard many times but has not been heard in a
particular construction, a language learner may infer that that particular verb-
construction pairing is not possible. Preemption takes into account both semantics
and frequency: when a verb has been heard in a particular construction, but not in
another related construction with similar meaning (e.g. the two locative
constructions), a language learner may infer that the verb cannon be used in that
second construction.
Stefanowitsch (2008) used adult grammaticality judgements of verbs in the
two dative constructions to test the entrenchment and preemption hypotheses. He
took corpus counts of 20 dative verbs that either appeared only in the double-object
dative construction (e.g. earn: Her books have earnt her a fortune/*Her books have
earnt a fortune to her) or the prepositional-object dative (e.g. explain: He explained
the procedure to them/*He explained them the procedure). The corpus counts were
used to calculate the degree of preemption and the degree of entrenchment
(Stefanowitsch referred to this as negative entrenchment) for each verb. Findings
showed a significant correlation between grammaticality judgements and degree of
entrenchment, but no significant correlation between grammaticality judgements and
degree of preemption. This finding contradicts that of Ambridge et al. (2014),
43
although the difference in methodology (correlation vs. mixed effects models) and
the corpora from which frequency counts were obtained may go some way to
explaining the different findings.
In an attempt to provide a more complete picture of the relative contributions
of entrenchment and preemption to children’s retreat from overgeneralisation errors,
Ambridge, Bidgood, Twomey, Pine, Rowland and Freudenthal (2015) tested the
predictions of the two hypotheses with several different alternations at the same time:
intransitive-transitive (The ball rolled/The man rolled the ball); dative (The boy gave
a present to the girl/The boy gave the girl a present); locative (The boy sprayed paint
onto the statue/The boy sprayed the statue with paint); and active/passive (The girl
kicked the boy/The boy was kicked by the girl). Children (aged 5-6 and 9-10 years)
and adults provided grammaticality judgements for sentences from all four of these
alternations. Results were in line with those of Stefanowitsch (2008): entrenchment
was a significant predictor of grammaticality ratings, whereas preemption was not.
However, Ambridge, Bidgood, Twomey, Pine, Rowland and Freudenthal (2015)
acknowledge that preemption does appear to play a role “when the preempting
construction is (a) particularly frequent relative to the error construction and (b)
particularly closely synonymous with the error” (p. 17). Thus, a specific preemption
effect is observed for the dative construction, for example (Ambridge et al., 2014). In
conclusion, Ambridge, Bidgood, Twomey, Pine, Rowland and Freudenthal (2015)
conclude that entrenchment and preemption should not be thought of as mechanisms
so much as effects of a process of competition between constructions, such as that
proposed under the FIT account.
When constructions are competing to convey the message of a boy causing a
girl to laugh, the competing constructions might be the intransitive-inchoative (The
girl laughed), the periphrastic causative (The boy made the girl laugh) and the
(ungrammatical) transitive-causative (*The boy laughed the girl). In avoiding
producing the ungrammatical transitive-causative construction, an entrenchment
effect may be observed as the speaker will have heard the verb used in the
intransitive-inchoative and periphrastic causative constructions many more times
than in the transitive-causative. This latter construction will therefore be activated
less than the other two. However, a preemption effect may also be observed: the
intransitive-inchoative construction, whilst much more frequent in the input than the
periphrastic causative, may not be the most relevant, as it does not allow the speaker
44
to express the causer of the laughing event. Thus, in order to avoid the
overgeneralisation error and still produce a sentence that expresses the entire
message, the periphrastic causative would have to win out over both the transitive-
causative and intransitive-inchoative alternatives. For verbs that appear only very
rarely in the periphrastic causative construction, the activation level may not be high
enough to block the production of the overgeneralisation error, *The boy laughed the
girl.
2.5 Summary
Evidence from the studies described in this chapter show some support for all three
of the main hypotheses of interest: Pinker’s (1989) semantic verb class hypothesis,
Braine and Brooks’s (1995) entrenchment hypothesis and, to a lesser extent,
Goldberg’s (1995) preemption hypothesis. However, what is clear is that no
individual hypothesis is capable of explaining all of the data. What is needed is an
approach that integrates both semantic and frequency information into account in its
explanation of children’s retreat from error, such as the FIT account (Ambridge &
Lieven, 2011; Ambridge et al., 2011).
The following three chapters present experimental evidence for an account
that integrates semantic and statistical information, using various different
methodological approaches. Chapter 3 examines the locative alternation (Christine
sprayed the flowers with water/sprayed water onto the flowers), using a novel-verb
grammaticality judgment study. Chapter 4 uses grammaticality judgments and error
elicitation tasks to examine the transitive-intransitive alternation (The ball
rolled/Robert rolled the ball). Finally, Chapter 5 goes beyond overgeneralisation
errors to examine how verb-in-construction semantic compatibility influences
children’s production of a construction that is known to be problematic until a
relatively advanced age: the passive. Chapter 6 concludes the thesis, summarising
how the evidence gathered in the experimental chapters fits the theories in question,
how this relates to nativist and constructivist approaches to language acquisition, and
suggesting next steps for research in this field.
45
Chapter 3: The retreat from locative overgeneralisation errors: A novel verb
grammaticality judgment study
3.0 Fit within the thesis
As discussed in Chapter 2, this thesis investigates three approaches to explaining
how children retreat from argument structure overgeneralisation errors: semantics
(Pinker, 1989), entrenchment (Braine & Brooks, 1995) and preemption (Goldberg,
1995). In addition, it will suggest how these three approaches might be integrated
into a single account, such as the FIT account (Ambridge & Lieven, 2011; Ambridge
et al., 2011). This chapter investigates the locative alternation: Lisa sprayed water
onto the roses (figure locative)/Lisa sprayed the roses with water (ground locative).
One reason for choosing the locative construction as the first to investigate in this
thesis is that children have been reported to make overgeneralisation errors in both
directions, both ground-only verbs used in the figure locative (e.g. I’m going to cover
myself with a screen *I’m going to cover a screen over me, Bowerman, 1982a)
and figure-only verbs used in the ground locative construction (e.g. I’m gonna pour
water onto it *I’m gonna pour it with water, Bowerman, 1981). For other often-
studied construction pairs, such as the dative (Bart gave Lisa the book/Bart gave the
book to Lisa), errors are usually only reported in one direction (e.g. Don't say that to
me *Don't say me that, Bowerman, 1978); in contrast, attested errors in the
locative alternation appear to be truly productive on both directions. A second reason
for using the locative is that, in terms of testing Pinker’s (1989) semantic verb class
hypothesis, there are relatively subtle differences between the subclasses. This means
that children might have difficulty realising which verbs are and are not able to
alternate. If they prove unable to do so, this would be evidence against the semantic
verb class hypothesis.
This chapter uses novel verbs to test the predictions of both the frequency and
semantic approaches, although the use of novel verbs is a particularly appropriate,
and stringent, way to test the importance of semantics. This is because, when
participants judge the grammaticality of sentences containing novel verbs, they have
no experience of hearing these verbs in sentences. They therefore have to judge the
grammaticality of sentences based on semantics, which they have inferred from
watching animations viewed during training. Novel verbs also allow us to use a
46
three-way contrast of verb frequency (high, low and novel) to test the statistical
mechanisms in question.
The study in this chapter finds support for both semantic and statistical
mechanisms in the retreat from overgeneralisation errors in the locative alternation.
However, the locative is a relatively rare construction. Chapter 4 therefore
investigates a much more frequent construction pair: the causative alternation (John
rolled the ball/The ball rolled). This alternation also allows us to distinguish between
the entrenchment and preemption hypotheses. Chapter 4 takes a different approach to
semantics, moving away from discrete semantic classes towards continuum of
semantic compatibility, more consistent with Ambridge and colleagues’ FIT account.
Alongside grammaticality judgments, Chapter 4 also moves on to investigate errors
in production, and takes a different approach to investigating semantics.
This chapter has been published as a paper in PLoS ONE (Bidgood et al.,
2014).
3.1 Introduction
As adults, we have the capacity for enormous creativity in language production: we
often produce utterances that we have never heard. To reach this stage, children must
acquire the grammar of the ambient language by forming generalisations about that
language from the input. However, children must also learn to restrict these
generalisations in order to avoid producing ungrammatical utterances (e.g. *I don’t
want it because I spilled it of orange juice [= I spilled orange juice onto my toast];
Bowerman, 1981).
Pinker (1989) listed various grammatical constructions that have two
alternating forms. The locative construction, for example, alternates between the
ground- (or container-) locative, as in The farmer loaded the wagon with hay, and the
figure- (or contents-) locative, as in The farmer loaded hay into the wagon. In the
first sentence, the wagon is most affected, as it changes state from empty to full. In
the second sentence, it is the hay that is most affected, as it is moved to a specific
location; the wagon may or may not end up full. Pinker (1989, p. 79) described this
change in how the event is construed as a “gestalt shift”. (For earlier work on these
constructions, see e.g. Hall, 1965; Fillmore, 1967; Anderson, 1971; Bowerman,
1982b; Levin & Rappaport, 1986).
47
When children hear verbs used in both the ground- and figure-locative
constructions (load, spray, stuff, etc.), they may create a generalisation that any verb
used in one of these constructions can also be used in the other, and this works well
for some verbs. A child hearing You splashed me with water, a ground-locative
construction, might generalise to the figure-locative construction to produce the
grammatical utterance, You splashed water onto me. However, some English verbs,
such as fill and cover, can only be used in the ground-locative construction (ground-
only verbs) and generalising these verbs to the figure-locative construction would
produce an ungrammatical utterance, such as *We filled toys into the box.
Conversely, some verbs, such as pour and spill, can only be used in the figure-
locative construction (figure-only verbs). Generalising these verbs to the ground-
locative construction would similarly produce overgeneralisation errors, such as
*Daddy poured my cup with juice.
One factor that could contribute to the retreat from overgeneralisation errors
is parental feedback: so-called ‘negative evidence’. It is undoubtedly the case that
some parents provide feedback on errors that their children make, either through
direct correction (e.g. C: *I filled mud into the hole, M: No, say “I filled the hole with
mud”) or implicitly, via rephrasing (e.g. M: That’s right, you filled the hole with
mud), facial expressions, misunderstandings or requests for clarification. Whilst
evidence suggests that such feedback is helpful (Chouinard & Clark, 2003), children
are unlikely to receive sufficient feedback of this type to account entirely for their
retreat from overgeneralisation errors, particularly for low frequency verbs.
Furthermore, some examples of parent-child interactions suggest that such feedback
may have only a limited effect on children’s language production (for reviews, see
e.g. Pinker, 1989, pp. 9-14; Marcus, 1993).
The current paper investigates the extent to which two mechanisms constitute
a solution to the ‘no negative evidence’ problem (Bowerman, 1988) and therefore
explain the retreat from overgeneralisation with locative constructions. The first of
these is Pinker’s (1989) semantic verb class hypothesis: while evidence exists in
support of this account, previous studies have primarily focussed on errors involving
the transitive-causative and dative constructions, which, for reasons outlined in the
following section, do not constitute as strong a test of the hypothesis. The second
mechanism is statistical learning, in the form of entrenchment (Braine & Brooks,
48
1995) or preemption (Goldberg, 1995). Again, the locative alternation is a
particularly good test of these hypotheses, as detailed below.
3.1.1 The semantic verb class hypothesis
Pinker’s (1989) semantic verb class hypothesis attempts to explain how children’s
developing knowledge of verb semantics could explain the retreat from
overgeneralisation errors. The proposed mechanism involves innate linking rules,
which link generic semantic structures (‘thematic cores’) to verb argument
structures: all verbs with the same thematic core are licensed in the same argument
structure. These groups of verbs are known as broad semantic classes.
Some verbs, such as spray and load, can appear in more than one argument
structure. Once children hear such examples, broad-range rules are formed (although
the set of possible alternations is constrained by the innate linking rules). These
allow verbs in related broad classes, such as figure and ground locative verbs, to
alternate between the two structures. Until this point in development, learning is
conservative and production is restricted to the use of verbs only in argument
structures already heard by the child.
Of course, not all verbs that are grammatical in one locative construction are
grammatical in the other, and this is due to idiosyncratic differences between verbs.
Pinker (1989, pages 273-4) proposed that, by replacing “each idiosyncratic piece of
information... with a parameter” and matching verbs on this more detailed level of
semantics, narrow semantic classes (or ‘subclasses’) are formed. It is only
membership in an alternating narrow class that enables a verb to be used
grammatically in the other argument structure, via a narrow-range rule.
According to the semantic verb class hypothesis, the cause of children’s
overgeneralisation errors is that children do not initially have well-developed
knowledge of verb semantics and do not necessarily know enough verbs in each
narrow class for these classes to have been accurately formed. Thus,
overgeneralisations occur as children occasionally apply the broad-range rule to
some verbs to which a narrow-range rule would not apply. There is some evidence
that children know that these productive forms are ungrammatical (1989, pp. 322-4).
Children retreat from error as the operation of narrow-range rules gradually
supersedes that of broad-range rules; the broad-range rules do remain in place,
49
however, and enable adults to produce ‘Haigspeak’ utterances (which the
speaker/writer again does not necessarily consider to be grammatical, 1989, pp. 152-
160).
Pinker (1989, pp. 126-7) specifies 15 narrow subclasses for locatives and
allocates each of the 146 verbs to one of these subclasses (with two exceptions, wrap
and string, which may each be the only members of their own respective subclasses).
The defining semantics of each subclass specify whether the verbs contained within
it can alternate between constructions, via a narrow-range rule, although even
alternating classes have a bias towards one of the two constructions. Table 3.1
(adapted from Ambridge, Pine & Rowland, 2012, p. 262, based on Pinker, 1989, pp.
126-7) details the 15 subclasses.
Figure-
(content-)
oriented
(into/onto
verbs)
Smear-type, Alternating (N=10),
designated reference category.
Simultaneous forceful contact and
motion of a mass against a surface
brush, dab, daub, plaster,
rub, slather, smear, smudge,
spread, streak
Stack-type, Alternating (N=3). Vertical arrangement on a
horizontal surface
heap, pile, stack
Spray-type, Alternating (N=7). Force is imparted to a mass, causing
ballistic motion in a specified
spatial distribution along a
trajectory
inject, spatter, splash,
splatter, spray, sprinkle,
squirt
Scatter-type, Alternating (N=4).
Mass is caused to move in a
widespread or nondirected
distribution
bestrew, scatter, sow, strew
Pour-type, Content-only (N=10).
A mass is enabled to move via the
force of gravity
dribble, drip, drizzle, dump,
ladle, pour, shake, slop,
slosh, spill
Coil-type, Content-only (N=6).
Flexible object extended in one
dimension is put around another
object (preposition is around)
coil, spin, twirl, twist, whirl,
wind
Spew-type, Content-only (N=8).
Mass is expelled from inside an
entity
emit, excrete, expectorate,
expel, exude, secrete, spew,
vomit
Glue-type, Content-only (N=9).
Verbs of attachment
attach, fasten, glue, nail,
paste, pin, staple, stick, tape
Ground-
(container-)
Stuff-type, Alternating (N=6). A
mass is forced into a container
against the limits of its capacity
cram, crowd, jam, pack,
stuff, wad
50
oriented
(with verbs)
Load-type, Alternating (N=3). A
mass of a size, shape, or type
defined by the intended use of a
container is put into the container,
enabling it to accomplish its
function
load, pack, stock
Fill-type, Container-only (N=21).
A layer completely covers a surface
bandage, blanket, coat,
cover, deluge, douse, edge,
encrust, face, fill, flood,
inlay, inundate, line,
occupy, pad, pave, plate,
shroud, smother, tile
Pollute-type, Container-only
(N=22). Addition of an object or
mass to a location causes an
aesthetic or qualitative, often
evaluative, change in the location
adorn, burden, clutter, deck,
dirty, embellish, emblazon,
endow, enrich, festoon,
garnish, imbue, infect, litter,
ornament, pollute,
replenish, season, soil,
stain, tint, trim
Soak-type, Container-only
(N=15). A mass is caused to be
coextensive with a solid or layer-
like medium
drench, impregnate, infuse,
interlace, interlard,
interleave, intersperse,
interweave, lard, ripple,
saturate, soak, stain,
suffuse, vein
Clog-type, Container-only (N=12).
An object or mass impedes the free
movement of, from, or through the
object in which it is put
block, choke, clog, dam,
plug, stop up, bind, chain,
entangle, lash, lasso, rope
Bombard-type, Container-only
(N=8). A set of objects is
distributed over a surface
bombard, blot, dapple,
riddle, speckle, splotch,
spot, stud
Alternating
verbs with
“unique
geometry”
that do not
fit into the
above
classes
(N=2)
Static of a linear object along a
surface
string
A flexible object conforms to part
of the shape of an object along two
or more orthogonal dimensions
wrap
Table 3.1. Pinker’s (1989) narrow-range subclasses for locative verbs, adapted from
Ambridge, Pine and Rowland (2012).
Further work has since been conducted aimed at defining the nature of the
verb classes more precisely (e.g. Boas, 2008; Levin, 1993). However, this work does
not change the basic prediction of the semantic verb class hypothesis to be tested
51
here, namely that children’s production of, and retreat from, overgeneralisation errors
will be predicted by their knowledge of the semantic class of the verb. In the present
study, all of the verbs chosen were classified in the same way by both Pinker (1989)
and Levin (1993), although it is worth noting that the organisation of verbs into
classes of this kind is not universally accepted (e.g. Braine & Brooks, 1995;
Goldberg, 1995; Ambridge, Pine & Rowland, 2012; Brinkmann, 1997; Fellbaum,
1990). It is also worth noting that the semantic verb class hypothesis cannot explain
verb frequency effects, which are also pervasive in the literature (as reviewed
below). Indeed, some authors (e.g. Stefanowitsch, 2008) have argued that apparent
semantic verb class effects are epiphenomenal, with learners acquiring verbs'
argument structure restrictions solely on the basis of surface-based statistical learning
mechanisms such as entrenchment and preemption. It is to these mechanisms that we
now turn.
3.1.2 The frequency hypothesis
Various accounts have attempted to explain how children are able to learn which
verbs can be used in which constructions based on statistical properties of the input
(e.g. Clark, 1988; Naigles & Hoff-Ginsberg, 1998). For example, the entrenchment
hypothesis (e.g. Braine & Brooks, 1995; Theakston, 2004; Ambridge et al., 2008)
proposes that, although children may be aware that it is possible to use certain verbs
in two alternating constructions, such as the ground- and figure-locative
constructions, they gradually learn that this is not the case for all verbs. While
children hear figure-only verbs, such as pour, frequently in their input, they never
hear them in the ground-locative construction. Eventually, this leads children to infer
that, if it were possible to use pour in this construction, they “would have heard it by
now”, and hence that ground-locative uses of this verb are ungrammatical for adult
speakers. An account that includes a related statistical mechanism (alongside a
semantic element) is preemption (e.g. Goldberg, 1995; 2006; 2011; Boyd &
Goldberg, 2011). This account proposes that only uses of the verb in a different
grammatical pattern that nevertheless yields the same meaning will lead to the
inference that the non-attested form is ungrammatical. For example, utterances such
as She poured water into the cup would preempt *She poured the cup with water, but
52
other semantically more distant uses (e.g., It's pouring with rain) would not (or, at
least, would do so to a lesser degree).
Ambridge, Pine and Rowland (2012) attempted to distinguish between the
effects of entrenchment and preemption on the retreat from overgeneralisation in the
locative construction, suggesting that both may play a role. However, their
entrenchment and preemption predictors were highly correlated, which made it
difficult to distinguish effects of one from the other (see also Boyd, Ackerman &
Kutas, 2012). For this reason, differentiating between entrenchment and preemption
is beyond the scope of the present study (see also e.g. Wonnacott, 2011, p. 2; Perfors,
Tenenbaum & Wonnacott, 2010, p. 612). For the remainder of this chapter, we will
therefore simply refer to the frequency hypothesis. Our findings and conclusions
could apply equally to the entrenchment and preemption hypotheses.
3.1.3 Existing evidence for the two accounts
Previous studies have provided evidence in support of both the semantic verb class
hypothesis and statistical learning accounts. However, these have primarily been
restricted to overgeneralisation errors relating to the causative alternation, such as
Homer broke the plate/The plate broke (e.g. Boyd et al., 2012; Ambridge et al.,
2009; 2011; Brooks & Tomasello, 1999; Brooks et al., 1999; Brooks & Zizak, 2002;
Naigles, Fowler & Helm, 1992; Naigles & Lehrer, 2002). While these studies
provide some support for both the semantic verb class hypothesis and the frequency
hypothesis, any successful account must be able to deal with all of the alternations
for which overgeneralisation errors are sometimes observed. Ambridge, Pine,
Rowland and Chang (2012) tested the predictions of the semantic verb class and
entrenchment hypotheses with the dative construction, finding support for both
theories, but only in their adult participants (see also Gropen, Pinker, Hollander,
Goldberg & Wilson, 1989, for support for broad and narrow verb classes in the
dative construction).
So, while the results of studies involving the causative alternation appear to
be consistent with both the semantic verb class and frequency hypotheses, both seem
to struggle in the domain of the dative alternation. One possible explanation is that
the dative is a special case, and that the semantic verb class and frequency
hypotheses can explain the retreat from overgeneralisation across a range of different
53
constructions. Another is that it is the causative alternation that is the special case,
with other constructions showing no semantic class or frequency effects. The aim of
the present paper is, thus, to test the scope of the two hypotheses by testing their
predictions against a third alternation: the locative.
3.1.4 The locative alternation
Like the dative, the locative alternation contains two relatively low frequency
constructions with fine-grained distinctions between the relevant narrow semantic
subclasses, and therefore constitutes a particularly good test case for both
hypotheses. It provides a strong test of the semantic verb class hypothesis because of
the sometimes very subtle differences between the narrow subclasses (see Table 3.1).
For example, with alternating spray-type verbs, a mass is caused to move via a force
imparted upon it, whereas, with ground-only pour-type verbs, a mass is simply
enabled to move via the force of gravity. In contrast, differences between subclasses
for the causative alternation seem more clear-cut: For example, verbs specifying the
manner of motion, such as bounce (The ball bounced / Bart bounced the ball),
alternate whereas verbs that specify the direction of motion, such as fall (The ball fell
/ *Bart fell the ball), do not (Pinker, 1989, pp. 130-4). In addition to the subtle
subclass distinctions in the locative alternation, for children to form the appropriate
subclasses, they would need to be able to observe the differences between them.
Again, this seems far less plausible for locative verbs than for causative verbs since,
in the locative example above, both the forces involved (e.g. gravity) and the subtle
difference between causing and enabling motion are difficult to observe.
Like the dative, the locative alternation also provides a strong test of the
frequency hypothesis due to the relatively low frequency of locative verbs,
particularly in comparison with verbs involved in the causative alternation. A paucity
of locative verbs (and, presumably, constructions) in the input could make it difficult
for statistical learning mechanisms to operate.
A further advantage of studying the locative construction, in this case over
both the causative and the dative constructions, is that it appears to be truly
productive in both directions. With regard to the dative alternation, all known errors
involve the overgeneralisation of prepositional-object (PO) verbs into the double-
object (DO) dative construction (e.g. Don't say that to me *Don't say me that,
54
Bowerman, 1978). We are aware of no reported cases of DO verbs being
overgeneralised into the PO construction (e.g. Homer bet Marge $10 *Homer bet
$10 to Marge). With regard to the causative alternation, the vast majority of errors
involve the overgeneralisation of intransitive-only verbs into the transitive-causative
construction (e.g. She cried *You cried her, Bowerman, 1982a). The converse
error, whilst attested (e.g. I didn't lose it *It won't lose, Lord, 1979), is extremely
rare. However, the locative is truly bidirectional, with many examples reported in the
literature of ground-only verbs being used in the figure locative (e.g. I’m going to
cover myself with a screen *I’m going to cover a screen over me, Bowerman,
1982b) and of figure-only verbs being used in the ground locative construction (e.g.
I’m gonna pour water onto it *I’m gonna pour it with water, Bowerman, 1981).
This bidirectionality of errors is a useful feature of the locative, because it
allows us to test for a possible confound: that children may be completing the
judgment task using task-based strategies, especially for novel verbs. For example, in
the causative study of Ambridge et al. (2008) and the dative study of Ambridge,
Pine, Rowland and Chang (2012), a task-based strategy of always rating intransitives
(in the former) or prepositional-object datives (in the latter) as acceptable would
yield adult-like judgments for these sentence types, since all were, in fact,
grammatical. Note that, in principle, children could quite easily establish such a
strategy on the basis of the high frequency, familiar verbs in the studies (e.g. Bart
laughed; Homer gave a book to Marge), and apply this strategy to lower frequency
and novel verbs.
Thus, of the three argument structure alternations studied with respect to the
problem of the retreat from overgeneralisation - in/transitive, dative and locative - the
latter constitutes the strongest test case for both the semantic verb class and
frequency hypotheses. It is therefore perhaps surprising that, of the three alternations,
the locative has received by far the least experimental attention. We are aware of
only three relevant studies: Gropen, Pinker, Hollander and Goldberg (1991a, b) and
Ambridge, Pine and Rowland (2012). Both Gropen et al. studies showed support for
Pinker’s broad semantic classes, and Ambridge, Pine and Rowland found some
support for both levels of semantic class, as well as frequency. However, Ambridge,
Pine and Rowland investigated the semantic verb class hypothesis using known
locative verbs; no novel verbs were included. Although the authors controlled for
attested usage by using verb frequency as a predictor in the regression analysis, for
55
familiar verbs, the extent to which participants are basing their ratings on semantics
alone, as opposed to attested usage, is difficult to ascertain.
3.1.5 The present study
The aim of the present study was to conduct a particularly strong test of the semantic
verb class and frequency hypotheses by (a) focussing on the locative alternation, and
(b) including both familiar and novel verbs. We obtained grammaticality judgment
data from children (aged 5-6 and 9-10) and adults for uses of high frequency, low
frequency and novel locative verbs (figure-only, ground-only and alternating) in both
locative constructions. We tested whether participants would be able to use verb
semantics to determine the grammaticality of sentences containing novel verbs, as
predicted by the semantic verb class hypothesis. We also tested whether participants’
tolerance of overgeneralisation errors when verbs are used in the inappropriate
construction decreased with each increasing level of verb frequency
(novel/low/high), as predicted by the frequency hypothesis.
A noteworthy aspect of this study is the fact that participants were taught
novel verbs, each of which had semantics consistent with only one of Pinker’s (1989)
narrow subclasses: two novel verbs each from a ground-only subclass, a figure-only
subclass and an alternating subclass. Participants’ ability to use the semantics of each
novel verb to make their grammaticality judgments is key to Pinker’s (1989)
proposal: without having the necessary subclasses in place, participants will be
unable to judge which locative construction is (un)grammatical for each novel verb.
3.2 Method
3.2.1 Ethics Statement
This study was approved by the University of Liverpool Ethics Committee. Informed
consent was obtained in writing both from adult participants and from the parents of
the children who took part.
56
3.2.2 Participants
The participants were 20 children aged 5-6 years (5;6-6;5. M=5;11), 20 children aged
9-10 years (9;6-10;5, M=9;11) and 20 adults aged 20-25 years. The children were
recruited from primary schools, and the adults from the University of Liverpool. All
participants were monolingual speakers of English, and had no known language
impairments.
3.2.3 Design and materials
3.2.3.1 Design
The experiment used a 3 x 2 x 3 x 3 x 2 mixed design. The between-subjects
variables were age of participant (5-6 years, 9-10 years, adult) and counterbalance
version (two groups based on which novel verb forms were paired with each
meaning). The within-subjects variables were semantic verb subclass (fill-type,
spray-type, pour-type; see below), verb frequency (high, low, novel) and sentence
type (ground-locative, figure-locative).
3.2.3.2 Test sentences and animations
Table 3.2 shows all verbs and test sentences used. Locative verbs were chosen based
on Pinker’s (1989) narrow subclasses (subsequently referred to simply as ‘classes’).
The first of these is the ground-only (or container-only) fill class in which “a layer
completely covers a surface”, the second is the figure-only (or contents-only) pour
class in which “a mass is enabled to move via the force of gravity”, and the third is
the alternating spray class in which “force is imparted to a mass, causing ballistic
motion in a specified direction along a trajectory”. For each class, two high
frequency and two low frequency verbs with similar semantics were chosen. (Mean
lemma frequency counts from the British National Corpus, 2007, are 5923 [range
750-18726] for high frequency verbs and 351 [range 111-658] for low frequency
verbs; see Table 3.2 for details.) Participants were also taught novel verbs with
similar meanings to the known verbs, two for each semantic class (see below for
details of the training method). The form-meaning pairings for novel verbs differed
57
for each counterbalance group in order to control for any effect of phonological
form.
Verb
Class
Frequ-
ency Verb
Sentence
Type Sentence
Fill
verbs
High
(18726)
Cover
*Figure *Bart covered mud onto Lisa
Ground Bart covered Lisa with mud
Low
(487)
Coat
*Figure *Bart coated mud onto Lisa
Ground Bart coated Lisa with mud
Novel
bredge/
blafe
*Figure *Bart bredged/blafed mud onto Lisa
Ground Bart bredged/blafed Lisa with mud
High
(10546)
Fill
*Figure *Lisa filled paper into the box
Ground Lisa filled the box with paper
Low
(111)
Line
*Figure *Lisa lined paper into the box
Ground Lisa lined the box with paper
Novel
chool/
tesh
*Figure *Lisa chooled/teshed paper into the box
Ground Lisa chooled/teshed the box with paper
Spray
verbs
High
(750)
Spray
Figure Lisa sprayed water onto the roses
Ground Lisa sprayed the roses with water
Low
(544)
Sprinkle
Figure Lisa sprinkled water onto the roses
Ground Lisa sprinkled the roses with water
Novel
tesh/
bredge
Figure Lisa teshed/bredged water onto the roses
Ground Lisa teshed/bredged the roses with water
High
(750)
Splash
Figure Homer splashed water onto Marge
Ground Homer splashed Marge with water
Low
(111)
Spatter
Figure Homer spattered water onto Marge
Ground Homer spattered Marge with water
Novel
dape/
nace
Figure Homer daped/naced water onto Marge
Ground Homer daped/naced Marge with water
Pour
verbs
High
(3461)
Pour
Figure Homer poured water into the cup
*Ground *Homer poured the cup with water
Low
(658)
Drip
Figure Homer dripped water into the cup
*Ground *Homer dripped the cup with water
Novel nace/
dape
Figure Homer naced/daped water into the cup
*Ground *Homer naced/daped the cup with water
High
(1306)
Spill
Figure Marge spilt juice onto the rug
*Ground *Marge spilt the rug with juice
Low
(195)
Dribble
Figure Marge dribbled juice onto the rug
*Ground *Marge dribbled the rug with juice
Novel
blafe/
chool
Figure Marge blafed/chooled juice onto the rug
*Ground *Marge blafed/chooled the rug with
juice
Table 3.2. All verbs and test sentences used in test trials. Verb frequency counts
(lemma counts from the British National Corpus, 2007) are provided in brackets.
58
For each of the verbs, a test sentence was created using each of the figure-
and ground-locative constructions (see Table 3.2). Thus, for each verb in the ground-
only fill class and the figure-only pour class, one sentence for each verb was
grammatical and one ungrammatical (e.g. *Lisa filled paper into the box; Lisa filled
the box with paper; Homer poured water into the cup; *Homer poured the cup with
water), whereas both sentences were grammatical for verbs in the alternating spray
class (e.g. Lisa sprayed the roses with water; Lisa sprayed water onto the roses).
Both sentences in each pair contained identical noun phrases.
For all test sentences, animations were created using Anime Studio Pro
Version 5.5 (2006) and presented to participants using a laptop computer.
Animations for both sentences in each test pair were identical, but each was
presented with the relevant pre-recorded test sentence. Animations served to ensure
that participants understood the intended meaning of the sentences, particularly those
including novel verbs. They also established the veracity of each of the descriptions,
thereby encouraging the participants, particularly the younger ones, to judge the
sentences on the basis of their grammaticality rather than their truth value.
3.2.3.3 Novel verb training sentences and animations
Each novel verb was assigned a meaning similar to, but subtly different from, its
semantic classmates in the study, whilst still being consistent with the class (e.g.
filling with a particular substance or pouring in a particular manner; see Table 3.2).
The English language includes verbs specifying both filling/coating with a particular
substance (e.g. to oil, to water, to paper) and pouring in a particular manner (e.g. to
dribble, to drip, to ladle). Thus, these novel verb meanings are neither non-language-
like in general nor non-English-like in particular.
For each novel verb, three animations were created in order to convey the
intended meanings to participants. For each of these animations, the novel verb was
given three times, always as a gerund. The sentences were as follows:
1. (before clip) Look what CHARACTER’s gonna do, it’s called VERBing.
2. (during clip) Look what CHARACTER’s doing, it’s called VERBing.
3. (after clip) So VERBing is... [followed by a brief definition, see Table 3.3].
59
The definitions were intended to clarify the meanings of each verb and point out the
important features of the action, which would enable learners to recognise each verb
as being consistent with the intended narrow semantic class. Importantly, novel verbs
were never presented in locative or transitive sentences during training (only as
simple intransitives), to prevent participants basing their judgments of the novel-verb
sentences on attested usage. Rather, according to the semantic verb class hypothesis,
learners should determine the locative construction(s) in which each verb can be used
on the basis of its semantics.
Novel verb Definition
Novel cover/coat like covering, except that it has to be with mud (like this)
Novel fill/line like filling, except that it has to be with paper (like this)
Novel
spray/sprinkle
like spraying, except that you have to press a button (like
this)
Novel
splash/spatter
like splashing, except that it has to be in big blobs (like this)
Novel pour/drip like pouring, except that it has to be in one big lump (like
this)
Novel spill/dribble like spilling, except that it has to be straight down in tiny
drops (like this)
Table 3.3. Novel verbs and definitions.
3.2.3.4 Grammaticality judgments
Participants rated sentences for grammatical acceptability using a five-point ‘smiley
face’ scale (see Figure 3.1 and Ambridge et al., 2008). The scale was presented with
no text or numbers. After viewing an animation and hearing the accompanying
sentence, children were asked to first choose a coloured counter, with green
indicating that the sentence ‘sounded good’ and red that it ‘sounded silly’. They then
placed the counter onto the scale to indicate how ‘good’ or ‘silly’ it sounded. The use
of counters was intended to enable younger children to indicate that they found a
sentence broadly acceptable or unacceptable, even if they were unable to provide a
more graded judgment (although this did not turn out to be the case). The
experimenter made a note of the judgment rating the child gave for each sentence.
Adults and older children were asked simply to tick one of the faces to provide their
judgment rating.
60
Figure 3.1. Five-point ‘smiley face’ scale for providing grammaticality judgments.
Participants were trained in the use of the judgment scale with a series of
seven training animations. The first four of these were designed to be clearly
acceptable or unacceptable, with the others designed to receive ratings somewhere in
between. Sentences were chosen based on ratings given by participants in previous
studies (see Table 3.4). Ratings for the first two sentences were given by the
experimenter, to demonstrate the use of the scale, and participants were given
feedback on their ratings for the five subsequent sentences. No feedback was given
during the experiment proper. Detailed descriptions of the training procedure are
given in Ambridge et al. (2008, pp. 106-7) and Ambridge (2011, pp. 122-3).
Sentence Typical score
The frog caught the fly 5
His teeth man the brushed 1
The cat drank the milk 5
The dog the ball played with 1
The man tumbled Bart into a hole 2 or 3
The magician vanished Bart 2 or 3
The funny clown giggled Bart 1 or 2
Table 3.4. Grammaticality judgment training sentences. ‘Sentences’ used in the
grammaticality judgment training trials, with their ‘typical’ scores (based on
Ambridge et al., 2008). The experimenter completed the first two trials to
demonstrate, with participants completing the remainder. Feedback was provided if
judgments were thought to be inappropriate.
3.2.4 Procedure
Participants were first taught the novel verbs and then received training on the use of
the grammaticality judgment scale (in both cases as described above). The main
study consisted of 36 test trials: one ground-locative sentence and one figure-locative
61
sentence using each of the six high frequency verbs, six low frequency verbs and six
novel verbs (see Table 3.2). Sentences were presented in a pseudo-random order,
such that two sentences containing the same verb were never given in succession. In
order to ensure that participants remembered the intended meaning of the novel
verbs, one of the training trials was repeated immediately before each test trial
containing a novel verb.
3.3 Results
Because the rating scale data are not true interval scale data, an empirical logit
transformation (Agresti, 2002) was applied. First, the rating scale was converted to a
proportion and were then transformed using the following formula: log (prop + 0.5) -
log(1 – [prop + 0.5]). All means and SEs are reported for raw scores. All post hoc
comparisons used Fisher’s Least Significant Difference tests. Data are available to
download from http://www.benambridge.com.
3.3.1 Preliminary analysis
A preliminary analysis, in the form of a 3x3x2 (age by verb class by sentence type)
mixed ANOVA, was performed on known verbs in order to confirm that the verb
type classifications (figure-only/ground-only/alternating) were correct for this group
of adult participants and that children were rating the sentences as expected.
Assuming that this is the case, the semantic verb class hypothesis predicts an
interaction of sentence type by verb class such that ground-locative uses are
preferred over figure-locative uses for verbs of the fill class with the reverse for verbs
of the pour class, and no preference for the spray class. This analysis, and all
subsequent analyses, were collapsed across the two counterbalance groups (which
differed only with regard to the pairings of phonological stem forms and novel verb
meanings), and across the two verbs in each cell of the design.
The ANOVA yielded several main effects. However, these will not be
discussed as they collapse across grammatical and ungrammatical sentences, and so
are not relevant to the hypotheses of the study. Importantly, as predicted, an
interaction of verb class by sentence type was observed (F(2, 114)=219.61, p<0.001,
ηp2=0.79). Analysis of this interaction revealed that, as predicted, for verbs in the fill
62
class, participants significantly preferred ground-locative uses (M=4.35, SE=0.05)
over figure-locative uses (M=3.16, SE=0.07, p<0.001). Conversely, for verbs in the
pour class, participants significantly preferred figure-locative uses (M=4.20,
SE=0.09) over ground-locative uses (M=2.43, SE=0.10, p<0.001). Also as expected,
for verbs in the alternating spray class, participants showed no preference for either
sentence type (ground M=4.18, SE=0.06; figure M=4.09, SE=0.07; p=0.12, n.s.).
A significant 3-way interaction of verb class by sentence type by age
(F(4,114)=9.05, p<0.001, ηp2=0.24; see Figure 3.2) indicated that the pattern of results
outlined above differed according to age group. This interaction was driven by the
fact that, whilst all age groups displayed the predicted pattern for the non-alternating
fill and pour verb classes, the adults also displayed an unexpected preference for
ground-locative uses of verbs from the alternating spray class, although a mean
rating of 4 or above still indicates that both sentence types were rated as broadly
acceptable. It is possible that this result reflects adults’ sensitivity to the holism
constraint: when an action has been wholly and successfully completed (as is the
case for the animations using alternating verbs in the present study), the ground-
locative construction is more felicitous that the figure-locative construction (cf. Lisa
taught the students French vs. Lisa taught French to the students). This is an issue to
which we will return in the Discussion.
Figure 3.2. Three-way interaction of age by verb class by sentence type for familiar
verbs.
63
3.3.2 Testing the semantic verb class hypothesis
In order to test the semantic verb class hypothesis, participants were taught six novel
verbs, two of which were semantically consistent with the ground-only fill class, two
with the alternating spray class and two with the figure-only pour class. Participants
were then asked to judge sentences containing each of these novel verbs for their
grammaticality. Each verb was presented in a figure-locative and a ground-locative
construction. The semantic verb class hypothesis predicts that, as with known verbs
of the same semantic classes, participants will judge figure-locative uses of the novel
fill verbs to be less acceptable than ground-locative uses of these verbs, with the
opposite pattern for the novel pour verbs, and no difference for the alternative uses of
the novel spray verbs.
These predictions were again tested by means of a 3x3x2 (age by verb class
by sentence type) mixed ANOVA, in this case conducted on the ratings for the novel
verbs only. As before, this analysis yielded several main effects, which will not be
discussed because they collapse across grammatical and ungrammatical sentences.
Importantly, as predicted, and in line with the results for all verbs, an interaction of
verb class by sentence type was observed (F(2, 114)=42.45, p<0.001, ηp2=0.43).
Analysis of this interaction revealed that, as predicted, for novel verbs in the ground-
only fill class, participants significantly preferred ground-locative uses (M=4.17,
SE=0.07) over figure-locative uses (M=3.52, SE=0.09, p<0.001). Conversely, and
again as predicted, for novel verbs in the figure-only pour class, participants
significantly preferred figure-locative uses (M=4.19, SE=0.08) over ground-locative
uses (M=3.18, SE=0.13, p<0.001). Unexpectedly, for novel verbs in the alternating
spray class, participants also showed a small but significant preference for ground-
locative uses (M=4.20, SE=0.10) over figure-locative uses (M=3.93, SE=0.10,
p=0.031), although a mean rating of around 4 or above still indicates that both
sentence types were rated as broadly acceptable. As previously noted, this may be
due to the holism constraint.
A significant 3-way interaction of class by sentence type by age (F(4,114)=4.27,
p=0.003, ηp2=0.13) indicated that the pattern of results outlined above differed
according to age group. As outlined in more detail below, this interaction was driven
by the fact that, whilst all groups displayed the predicted pattern for the novel verbs
in the non-alternating pour class, only older children and adults showed the expected
64
preference for ground-locative uses of novel verbs in the non-alternating fill class,
and only the adults displayed the unexpected preference for ground uses of novel
verbs from the alternating spray class (see Figure 3.3).
Figure 3.3. Three-way interaction of age by verb class by sentence type for novel
verbs.
As predicted by the semantic verb class hypothesis, the 5-year-olds showed
no significant preference for novel alternating spray class verbs in figure-locative
uses (M=3.65, SE=0.19) or ground-locative uses (M=3.78, SE=0.20, p=0.82, n.s.).
Also as predicted, they significantly preferred figure-only pour verbs in figure-
locative uses (M=3.78, SE=0.14) over ground-locative uses (M=3.05, SE=0.21,
p=0.008). These results suggest that they have identified the verb classes of these
novel verbs correctly, and are using this information to judge the grammaticality of
the verbs’ use in the alternative locative constructions. Contrary to the prediction,
however, the 5-year-olds displayed no significant preference for novel ground-only
fill class verbs in ground-locative uses (M=3.53, SE=0.12) over figure-locative uses
(M=3.48, SE=0.16, p=0.74, n.s.). It is possible that this youngest group of children
had not fully grasped the complex semantics of fill class verbs, which may be more
complex than those of the pour class (see Gropen et al., 1991b, and Introduction,
above).
65
The results for the 9-year-olds are all as predicted by the semantic verb class
hypothesis: no preference for novel alternating spray class verbs in either figure-
locative uses (M=4.18, SE=0.16) or ground-locative uses (M=4.25, SE=0.17, p=0.69,
n.s.), a significant preference for figure-only pour class verbs in figure-locative uses
(M=4.23, SE=0.14) over ground-locative uses (M=3.23, SE=0.16, p<0.001), and a
significant preference for ground-only fill class verbs in ground-locative uses
(M=4.18, SE=0.14) over figure-locative uses (M=3.55, SE=0.10, p<0.001).
Adults also displayed the predicted preferences for the novel figure-only pour
class verbs and the novel ground-only fill class verbs. They preferred figure-only
pour class in figure-locative uses (M=4.58, SE=0.14) over ground-locative uses
(M=3.25, SE=0.28, p<0.001) and they preferred novel ground-only fill class verbs in
ground-locative uses (M=4.80, SE=0.08) over figure-locative uses (M=3.53,
SE=0.18, p<0.001). Both of these results are in line with the predictions of the
semantic verb class hypothesis. Unexpectedly, however, the adult participants also
preferred the novel alternating spray class verbs in ground-locatives (M=4.65,
SE=0.16) over figure-locatives (M=3.95, SE=0.20, p=0.002). This unexpected result
parallels the findings observed for adults with familiar verbs, and may again be
explained by the holism constraint (see Discussion). The fact that the 9-year-olds did
not show this preference, whilst otherwise displaying an adult-like pattern of results,
indicates that the holism constraint (as applied to the ground-locative construction, at
least) may not be fully acquired until very late in development.
3.3.3 Testing the frequency hypothesis
To test the frequency hypothesis, we calculated difference scores for grammaticality
judgment ratings for ‘grammatical’ sentences (ground-locative uses of fill class
verbs; figure-locative uses of pour class verbs) minus ‘ungrammatical’ sentences
(figure-locative uses of fill class verbs; ground-locative uses of pour class verbs) for
high frequency, low frequency and novel verbs in both of these non-alternating
classes. These difference scores represent the degree of preference for grammatical
over ungrammatical verb uses (or, perhaps more importantly for our purposes, the
degree of dispreference for ungrammatical verb uses relative to matched
grammatical alternatives). Alternating verbs were not included in this analysis since
the frequency hypothesis only makes predictions regarding the degree of
66
unacceptability of ungrammatical verb uses (for alternating verbs, by definition,
neither figure- nor ground-locative uses are ungrammatical).
The frequency hypothesis predicts that the largest difference scores will be
observed for the high frequency verbs, smaller difference scores for the low
frequency verbs and the smallest difference scores for the novel verbs. That is,
increased exposure to a verb in grammatical sentences is predicted to increase the
strength of the inference that non-attested uses are not permitted, and hence the
extent to which participants will rate ungrammatical uses of that verb as
unacceptable.
A 3x2x3 (age by verb class by verb frequency) ANOVA revealed that all
three main effects were significant. The main effect of verb class (F(1,57)=29.83,
p<0.001, ηp2=0.34) indicates that participants showed a larger dispreference for
ungrammatical uses of pour class verbs (M=1.52, SE=0.10) than fill class verbs
(M=1.01, SE=0.06). While the frequency hypothesis makes no predictions about verb
class, this result is consistent with the results of the semantic verb class analysis,
which found that participants were less tolerant of overgeneralisation errors with
novel fill-type verbs than novel pour-type verbs.
The main effect of age (F(2,57)=18.08, p<0.001, ηp2=0.39) demonstrates that
adults (M=1.78, SE=0.12) showed a greater degree of dispreference for
ungrammatical sentences than both 9-year-olds (M=1.24, SE=0.08) and 5-year-olds
(M=0.79, SE=0.14), and that 9-year-olds showed a greater degree of dispreference
for such uses than 5-year-olds (all comparisons were significant at p<0.01 or better).
This result could be interpreted as showing support for the frequency hypothesis, as
adults will have had more exposure to grammatical uses of the relevant verbs than 9-
year-olds who, in turn, will have had more exposure than 5-year-olds. For this
interpretation to be correct, the important factor would have to be absolute frequency
of exposure to the verbs in competing constructions (e.g. total number of ground-
locative uses of fill), which obviously increases with age, as opposed to relative
frequency (e.g. proportion of uses of fill in the ground-locative construction as
opposed to other constructions), which presumably stays relatively constant across
development. However, the lack of interaction between age and verb frequency (see
below) suggests that this is not the case. That is, adults did not display a larger
frequency effect (i.e. larger between-verb differences) than children, which one
would expect if the relevant factor were absolute differences in verb frequency. It is
67
therefore likely that the main effect of age was simply due to older participants
performing better on the task.
Importantly, as predicted by the frequency hypothesis, a main effect of verb
frequency was observed (F(2,114)=38.25, p<0.001, ηp2=0.40; Figure 3.4) such that
participants showed a greater dispreference for ungrammatical uses of the high
frequency verbs (M=1.87, SE=0.11) than either the low frequency verbs (M=1.10, SE
= 0.09, p<0.001) or the novel verbs (M=0.83, SE=0.10, p<0.001), which also differed
significantly from each other in the predicted direction (p=0.050), although this last
difference was much smaller.
Figure 3.4. Main effect of verb frequency.
The analysis revealed no significant interactions of frequency by age
(F(4,114)=0.17, p=0.96, n.s., ηp2=0.01), verb class by age,(F(2,57)=1.74, p=0.19, n.s.,
ηp2=0.06), verb class by frequency (F(2,114)=1.84, p=0.16, n.s.,, ηp
2=0.03) or
frequency by verb class by age (F(4,114)=0.94 , p=0.45, n.s., ηp2=0.03).
3.4 Discussion
The aim of the present study was to conduct a particularly strong test of the semantic
verb class hypothesis (Pinker, 1989) and the frequency hypothesis (e.g. Braine &
68
Brooks, 1995; Goldberg, 1995) by (a) focussing on the locative alternation, and (b)
including both familiar and novel verbs. To this end, we obtained, from children
(aged 5-6 and 9-10 years) and adults, judgments of figure- and ground-locative
sentences containing high frequency, low frequency and novel verbs consistent with
figure-only, ground-only and alternating narrow semantic classes.
The findings suggest that, in general, participants were able to use the
semantics of each novel verb to align them with the ground-only fill class, the
alternating spray class or the figure-only pour class, although the youngest group of
children were unable to do so for novel fill-type verbs, and adults showed an
unexpected preference for ground-locative uses of novel spray-type verbs. The
findings of the present study also provide support for the frequency hypothesis:
participants in all age groups displayed a greater dispreference for overgeneralisation
errors with high frequency than with low frequency familiar verbs, and for errors
with both of these groups than with novel verbs.
3.4.1 The role of semantics
According to Pinker’s (1989) semantic verb class hypothesis, locative verbs fall into
one of two broad semantic classes. A broad-range rule links entries for alternating
verbs such as spray, which appear in both broad classes, allowing verbs attested in
one locative construction to be used in the other (e.g. Lisa sprayed the flowers with
water Lisa sprayed water onto the flowers). Overgeneralisation errors occur when
this rule is incorrectly applied to non-alternating verbs, such as fill and pour, and
cease only when children acquire the more specific narrow semantic subclasses and
narrow-range rules that allow the alternation to be restricted to verbs whose
semantics are compatible with the core meanings of both locative constructions.
The main test of Pinker’s hypothesis in the current study involved novel
verbs. Participants were taught six novel verbs with semantics consistent with one of
Pinker’s narrow subclasses of locative verbs: two each were consistent with (a) the
ground-only fill class, (b) the figure-only pour class, and (c) the alternating spray
class. Participants provided grammaticality judgments for ground-locative and
figure-locative uses of each of the novel verbs with results showing that, as
predicted, participants judged ground-locative uses of novel fill-type verbs to be
significantly more acceptable than figure-locative uses of these verbs, with the
69
opposite pattern observed for novel pour-type verbs. Since these verbs were never
presented in locative constructions during training, participants must have been using
verb semantics, as opposed to attested usage, to make these judgments. The subtle
differences between subclasses of locative verbs, which are also not easily
observable, make the locative alternation a particularly strong test of the semantic
verb class hypothesis. In addition, the fact that both some figure-locative and some
ground-locative sentences were ungrammatical allows us to rule out the possibility
that participants were using a task-based strategy to identify the ungrammatical
sentences (cf. Ambridge et al., 2008; Ambridge, Pine, Rowland & Chang, 2012).
Thus, the results of this study clearly point to an important role for verb semantics in
the retreat from overgeneralisation errors in the locative construction.
The semantic verb class hypothesis predicts no preference for either locative
construction for alternating spray-type verbs. However, while both constructions
were judged to be broadly grammatical, adult participants demonstrated an
unexpected preference for ground-locative uses of both familiar verbs and novel
verbs conforming to the semantics of this subclass. Therefore, one possibility is that
adults simply have a general preference for the ground-locative construction for
alternating verbs (although this is inconsistent with a strict interpretation of Pinker
[1989, p. 127], who lists spray-type verbs as being “content-oriented”, such that any
preference involving these alternating verbs should have been for the figure-locative
construction).
A possible explanation for the unexpected preference for ground-locative
uses of alternating verbs can be found in the holism constraint. This constraint
applies to ground-only locative verbs such as fill and cover, where the object must be
completely filled or covered, respectively, in order for the sentence to be an accurate
description of the event. The constraint also applies to the ground-locative
construction itself: one semantic feature of this construction, but not the figure-
locative construction, is that the ‘ground’ (e.g. the container) must be wholly
affected. Indeed, it is the incompatibility of the semantics of the figure-locative
construction and the semantics of verbs such as fill and cover that makes figure-
locative sentences using these verbs ungrammatical.
It is possible that participants may have preferred the ground-locative uses of
alternating verbs included in this study because, in the training for the novel verbs
and all test animations, the ‘location’ or ‘ground’ was always completely affected
70
(e.g. water splashed onto all of it). It was necessary to create the animations in this
way in order to keep the same methodology across all verbs and classes, since,
without being completely splashed with water, the animation would have been
inconsistent with the ground-locative construction. The animations could therefore
be considered to be more consistent with the semantics of the ground-locative
construction than with those of the figure-locative construction. The results also
suggest a developing knowledge of the holism requirement, as applied to individual
verbs, between the age of 5 and adulthood, which in turn provides further support for
the semantic verb class hypothesis. Unlike the older children and adults, the 5- to 6-
year-olds preferred figure-locative uses of novel pour-type verbs but showed no
preference for either argument structure for novel fill-type verbs. This suggests that
these children were unable to appreciate the holism requirement of the novel fill-type
verbs they were taught based on the animations they viewed during training (see also
Gropen et al., 1991a). The disparity between young children’s judgment data with
novel and familiar verbs may also indicate that these children are basing their
grammaticality judgments with familiar verbs on attested usage as opposed to, or in
addition to, verb semantics.
Additional support for the importance of a developing knowledge of the
holism constraint, as applied to the ground-locative construction, is the fact that only
the adult participants gave different judgment scores for the two locative uses of
alternating spray-type verbs (for both known and novel verbs), although both
constructions were judged to be broadly grammatical. This indicates knowledge of
the importance of context to the semantics of the alternative locative constructions
themselves, which may not yet have developed in the children we tested, leading
adults to judge ground-locative uses of spray-type verbs as more acceptable than
figure-locative uses of these verbs, based on the animations they viewed.
3.4.2 The role of frequency
The frequency hypothesis (e.g. Braine & Brooks, 1995; Goldberg, 1995) emphasises
the importance of the statistical properties of the input in children’s language
acquisition. Under this hypothesis, children retreat from overgeneralisation errors by
inferring, from their absence in the input, that certain argument structures cannot be
used with certain verbs. The more a child hears, for example, the verb fill used in
71
different constructions with a similar meaning (preemption) or a different
construction of any kind (entrenchment) without also hearing it in the figure-locative
construction, the better able they are to determine that it is not possible to use fill in
the latter. This hypothesis therefore predicts that participants will judge
overgeneralisation errors with high frequency verbs to be less acceptable than
equivalent overgeneralisation errors with low frequency verbs.
Results from the current study provide support for the frequency hypothesis.
Participants of all ages showed the same patterns of dispreference for
overgeneralisation errors, with higher dispreference scores for such errors with high
frequency verbs, lower scores for low frequency verbs, and the lowest dispreference
scores for novel verbs, which essentially have a frequency of zero in the input. This
finding replicates that of Ambridge, Pine and Rowland (2012), who found a negative
correlation between verb frequency and the acceptability of errors across a wider
range of locative verbs. So, despite the low frequency of locative verbs and
constructions in the input, the effects of this mechanism can clearly be seen in all age
groups tested here.
The frequency hypothesis could be interpreted in two ways: either absolute
frequency of a verb or the relative frequency of that verb in competing constructions
could be taken as the important factor in the retreat from overgeneralisation. Initially,
the finding that participants’ dispreference for overgeneralisation errors increased
with age appears to show support for the interpretation favouring absolute frequency,
since the absolute frequencies of the relevant verbs in different constructions will
increase with age, whilst the relative frequencies are likely to remain fairly constant
throughout development. However, the fact that no interaction between age and verb
frequency was observed counts against this interpretation. Provided that the ratio of
high to low frequency verbs in the input remains relatively stable for all ages, an
absolute frequency interpretation of the frequency hypothesis would have predicted
an increasing difference in dispreference scores for overgeneralisation errors between
verbs of different frequencies as the age of participants increased. The main effect of
age observed here is therefore likely to be due to older participants simply
performing better on the task. So, whilst the present study did not specifically
investigate this aspect of the frequency hypothesis, findings suggest that the relative
frequency of a verb in competing constructions might be the most important
statistical factor in the retreat from overgeneralisation.
72
3.4.3 Explaining the retreat from overgeneralisation
The predictions of both the semantic verb class hypothesis and the frequency
hypothesis have been supported by the findings of the current study: semantics and
statistics clearly both have a role to play in the retreat from overgeneralisation.
However, neither of these accounts in its current form can explain both the frequency
effect and the fact that participants were able to provide grammaticality judgments
for novel verbs in line with those of semantically-related familiar verbs. In order to
explain the retreat from overgeneralisation errors more fully, an account must be
posited that can explain both of these effects, such as Perfors et al.’s Bayesian
account (2010) or Ambridge et al.’s FIT account (Ambridge et al., 2011; Ambridge
& Lieven, 2011) (see also Alishahi & Stevenson, 2008; Chang, Dell & Bock, 2006;
MacWhinney, 2004; Tomasello, 2003).
This study has shown that, as predicted by the semantic verb class hypothesis,
children and adults are able to use the semantics of novel verbs to judge their
grammaticality in locative sentences in line with verbs with similar semantics. As
predicted by statistical learning accounts, children and adults judge errors with high
frequency verbs to be worse (in comparison with their grammatical counterparts)
than errors with low frequency verbs, which in turn are judged to be worse than
errors with novel verbs. Thus, this paper adds to previous research indicating the
importance of both semantics and statistics in children’s retreat from
overgeneralisation errors, and in language acquisition more widely. Future empirical
and computational work should focus on testing accounts, such as those mentioned
here, that integrate both of these mechanisms.
3.5 Acknowledgements
We would like to thank the schools, teachers, parents and children who made this
research possible, as well as the undergraduate project students who assisted with
data collection.
73
Chapter 4: How do children retreat from overgeneralisation error? Evidence
from the causative alternation
4.0 Fit within the thesis
The study in Chapter 3 investigated the locative alternation. It found support for both
semantic and statistical mechanisms. However, the locative is a relatively rare
construction. This chapter investigates the causative alternation (John rolled the
ball/The ball rolled). This alternation is an important test of the hypotheses in
question due to its frequency and the higher frequency of overgeneralisation errors
reported in these sentences types (see Pinker, 1989, pp. 22-25). So, whereas Chapter
3 provided a critical test of the semantics hypothesis in particular, this chapter
investigates the ability of semantic and statistical approaches to explain the errors
children make most frequently. Using a wider range of verbs, with a wider range of
frequencies, this study also allows us to distinguish between the entrenchment and
preemption hypotheses and to assess the relative contribution of each to the retreat
from overgeneralisation. This chapter takes a different approach to semantics to that
of Chapter 3. Rather than using discrete semantic classes, we created continuous
semantic variables from semantic ratings provided by a group of adult participants.
This approach allows more flexibility to investigate widely-observed lexical effects,
as outlined in Chapter 1.
This chapter also takes a multi-method approach. As well as using the
grammaticality judgment paradigm of Chapter 3, it uses production-priming to
investigate the errors that children actually produce. Although it is, in principle,
possible to investigate children’s production of errors through the use of corpora and
diary studies, the paucity of overgeneralisation errors in corpus studies and the
potential for observer bias in noticing errors in diary studies means that elicitation in
an experimental setting is preferable. The large number of errors that children made
in the production study (Experiment 3, below) allowed us to use powerful statistical
techniques to test the relative contribution to children’s retreat from
overgeneralisation errors of semantics, entrenchment and preemption in the same
model.
As in Chapter 3, the experiments reported here find support for both semantic
and statistical mechanisms. However, the way in which the semantics mechanism is
74
operating is not always clear, with some statistical predictions being in the opposite
direction to our expectations. Possible reasons for this are discussed. While support
for entrenchment is strong and fairly consistent, the picture for preemption is far less
clear. In fact, preemption seems to operate as expected in the production study only.
Taken together with the findings of Chapter 3, then, both statistical and semantic
mechanisms appear to play a role in children’s retreat from overgeneralisation errors,
although further study is required to narrow down the relative contributions of these
mechanisms in different alternations. Chapter 5 follows on from Chapters 3 and 4 by
investigating the role of semantics in a wider context. It uses the priming
methodology of the current chapter to examine children’s acquisition of the passive.
It also uses two approaches to semantics: the traditional, class-based approach taken
in the previous chapter, and the fine-grained semantic continuum approach taken in
the current chapter.
This chapter is currently being prepared for submission to a peer-reviewed
journal.
4.1 Introduction
In order to attain adult-like levels of productivity in language, children must be able
to generalise patterns they have observed in adults’ speech to new forms. For
example, a child might hear pairs of sentences such as The plate broke and Homer
broke the plate, and The window opened and Marge opened the window. These pairs
are examples of the causative alternation, in which verbs are used in both
intransitive-inchoative sentences (with no external agent expressed) and transitive-
causative sentences (with an explicit causative agent). Over time, the child will hear
this pattern repeated with many verbs. On hearing an intransitive sentence containing
a new verb, such as The ball bounced, they might generalise the pattern they have
learned and produce the novel (for them) sentence, Lisa bounced the ball.
However, while generalisation is key to language development,
overgeneralisations can also result. The same child who has created the sentence
Lisa bounced the ball to enable them express the causer of the ball’s bouncing might
also want to express the causer of their own giggling, leading to an ungrammatical
sentence like *You giggled me. Such errors have been reported in various studies,
perhaps most notably in Bowerman’s diary studies, e.g. *You just cried me (1981),
75
*Do you want to see us disappear our heads? (1988). To avoid making
transitivisation errors such as these, the child must learn to restrict the application of
the causative alternation they have observed to only the appropriate verbs.
Three main approaches to solving this problem currently exist (explained in
detail below), with varying levels of empirical support: the entrenchment hypothesis
(Braine & Brooks, 1995); the preemption hypothesis (e.g. Goldberg, 1995); and the
semantic verb class hypothesis (e.g. Pinker, 1989). Most of the empirical support for
these hypotheses comes from grammaticality judgment studies, in which participants
rate how (un)grammatical particular verbs sound in particular constructions. These
studies are useful, but give us little information about children’s production of
overgeneralisation errors. While production data do exist, diary studies (e.g.
Bowerman, 1982a; Lord, 1979) are, by their nature, limited in scope, and
experimental production studies (e.g. Brooks & Tomasello, 1999; Boyd et al., 2012)
have thus far tended to rely on a small number of novel verbs.
The current paper answers questions raised by the generalisability of current
data by using a multi-methodological approach. Firstly, grammaticality judgments
ensure that the current findings are in line with several recent studies of different
alternating constructions. We then use production-priming to elicit
overgeneralisation errors from young children, using a large number of real verbs, to
investigate how these errors pattern as a function of distributional and semantic
predictors. Our production data allow us to examine the validity of conclusions
drawn on the basis of grammaticality judgment data alone, both in the current paper
and in previous studies.
4.1.1 Entrenchment
Some proponents of statistical learning approaches have suggested that distributional
information in the input can help children avoid and/or retreat from
overgeneralisation errors. The entrenchment hypothesis (Braine & Brooks, 1995)
proposes an inference-from-absence mechanism. For example, a child will hear the
verb giggle used in various sentence structures (examples from CHILDES
[MacWhinney, 2000], Thomas corpus [Lieven, Salomo & Tomasello, 2009]): You
always get hiccups when you giggle; Why are you giggling?; That just makes you
giggle, doesn’t it?; You’re cheeky giggling away there; etc. The more often a child
76
(or adult) hears a verb without hearing it in the alternative construction, the more
information they have on which to make the inference that the unattested form must
be ungrammatical. Therefore, the entrenchment hypothesis predicts that
grammaticality judgments will vary with verb frequency: the higher the frequency of
the verb, the worse the ungrammatical, overgeneralised sentence will be judged to
be. Similarly, this hypothesis predicts that children will produce fewer errors with
verbs of higher than lower frequency.
Theakston (2004) investigated the role of entrenchment in the retreat from
overgeneralisation errors using a grammaticality judgment paradigm. Adults and
children (aged 5 and 8 years) heard sentences containing argument structure
overgeneralisation errors of different types (e.g. *She came me to school [cf. I came
to school], *I poured you with water [cf. I poured water onto you]). Half of the verbs
were high-frequency (e.g. come, pour) and half were low-frequency equivalents of
the same semantic class (as classified by Levin, 1993) (e.g. arrive, dribble). Children
indicated a binary grammatical/ungrammatical judgment, whereas adults’ judgments
were given on a 7-point scale. Results showed that participants of all ages judged the
overgeneralised sentences containing high-frequency verbs to be significantly more
ungrammatical than their equivalents containing low-frequency verbs. Thus, this
study provided strong support for the entrenchment hypothesis.
In the current paper, we test the entrenchment hypothesis using corpus counts
of verb frequency in all constructions. As verb frequency increases, we predict (a)
decreasing acceptability of ungrammatical, overgeneralised sentences in our
grammaticality judgment tasks, and (b) a decreasing error rate in our production-
priming study.
4.1.2 Preemption
A related statistical-learning mechanism is preemption (e.g. Goldberg, 1995),
although this hypothesis also involves an element of semantics. According to this
approach, hearing a verb used only in constructions with similar meanings will lead
to the inference that the unattested form is ungrammatical. In the case of the
transitive-causative, the construction with closest meaning is the periphrastic
causative: X made Y VERB. So, according to the preemption hypothesis, hearing a
verb such as giggle in periphrastic causative sentences like Bart made Maggie giggle
77
(but never transitive-causative sentences like *Bart giggled Maggie) should lead
children to realise that this is the construction they must use if they want to express
the causer of the giggling. The more often a child (or adult) hears a verb in the
competing, preempting construction, the more information they have on which to
make the inference that the unattested form must be ungrammatical. Therefore, the
preemption hypothesis predicts that grammaticality judgments will vary with verb
frequency in the preempting construction: the higher the frequency of the verb in that
construction (e.g. the periphrastic causative), the worse the ungrammatical,
overgeneralised sentence (e.g. the transitive-causative) will be judged to be.
Similarly, this hypothesis predicts that children will produce fewer errors with verbs
of higher frequency in the preempting construction than those with a lower frequency
in that construction.
Brooks and Zizak (2002) tested the predictions of the preemption hypothesis
in an elicited production study using novel verbs, with children aged 4 and 6-7 years.
The use of novel verbs allows for strict control of the number of exposures
participants receive to both preempting and non-preempting input sentences. All
children were taught two novel verbs (dack and tam), one of which was heard in 36
transitive sentences (e.g. The rabbit is dacking the car), the other in 36 intransitive
sentences (e.g. The house is tamming). Children were split into three groups. In the
No Preemption group, only the transitive and intransitive sentences were heard, with
no preempting alternatives. In the Alternative Construction group, prempting
sentences were also presentented, with the periphrastic causative (e.g. The rabbit
made the house tam) designed to preempt an ‘ungrammatical’ transitive sentence
(e.g. *The rabbit tammed the house), and the passive (e.g. The car is getting dacked)
designed to preempt an ‘ungrammatical’ intransitive sentence (e.g.*The car dacked).
In the English Suppletive group, along with the transitive and intransitive training
sentences, children heard sentences using known verbs (e.g. The car is swinging, The
rabbit bounced the house), which provided an alternative, grammatical way to
describe the action in the alternative transitivity. At test, the older children (but not
the younger ones) in the Alternative Construction group were significantly less likely
to produce a sentence violating the assigned transitivity of the novel verbs than the
children in either of the other groups (which did not differ significantly from each
other). Thus, support for the preemption hypothesis was found for these slightly
older children.
78
In the current paper, we test the preemption hypothesis using corpus counts of
verb frequency in preempting constructions. Following Brooks and Zizak (2002; see
also Brooks & Tomasello, 1999), we use the periphrastic causative (e.g. Homer made
the fish swim) as the preempting construction for transitivisation errors (e.g. The fish
swam *Homer swam the fish) and the passive (e.g. The ball was kicked) as the
preempting construction for intransitivisation errors (e.g. Homer kicked the ball
*The ball kicked). As verb frequency in the appropriate preempting construction
increases, we predict (a) decreasing acceptability of ungrammatical, overgeneralised
sentences in our grammaticality judgment tasks, and (b) a decreasing error rate in our
production-priming study.
4.1.3 Semantics
The final hypothesis under investigation in the current paper is the semantic
verb class hypothesis (e.g. Pinker, 1989; Levin, 1993). Under this approach, the
semantics of some verbs allow them to alternate between the transitive-causative
construction and the intransitive-inchoative construction, whereas the semantics of
other verbs mean that they are compatible with only one or other of these
constructions. As children’s knowledge of verb semantics is refined, they are able to
avoid overgeneralisation errors. In Pinker’s original proposal (1989), verbs are
grouped into ‘semantic classes’. Members of each class have related semantics and
behave in the same way in terms of their permissible argument structures. So, for
example, ‘verbs of extrinsic change of physical state’ (p. 130) such as open, melt and
shatter can alternate between intransitive and transitive sentences: The door
opened/Marge opened the door. In contrast, ‘verbs of emotional expression’ (pp.
130-1) such as cry, smile and blink are all intransitive-only verbs: The girl
smiled/*The joke smiled the girl. More recent approaches (e.g. Boas, 2008;
Fellbaum, 1990; Levin, 1993; van Valin, 2005) have offered slightly different
explanations of the organisation of verbs in terms of their semantics.
Gropen et al. (1991a) tested Pinker’s (1989) semantic verb class hypothesis
using the case of the locative alternation. In this alternation, some verbs are able to
alternate between the ground locative (e.g. Lisa sprayed the flowers with water) and
the figure locative (e.g. Lisa sprayed water onto the flowers) constructions. Others
are only grammatical in the ground locative (e.g. Lisa filled the cup with water vs.
79
*Lisa filled water into the cup) or the figure locative (e.g. Lisa poured water into the
cup vs. *Lisa poured the cup with water). ‘End-state verbs’ such as fill describe a
change of state of the container (here, the cup, which becomes full), whereas
‘manner verbs’ like pour describe the manner of motion of the contents (here, the
water, which moves downwards in a steady stream). Gropen et al. taught two novel
verbs (keat and pilk), one each from end-state and manner verb classes, to adult and
children (aged 3, 5 and 7 years), then elicited sentences containing these verbs to test
if participants were able to use verb semantics to produce ‘grammatical’ sentences, in
line with known verbs of the same semantic class. At test, participants produced
significantly more figure locative responses with novel manner verbs than novel end-
state verbs, indicating that participants were able to use verb semantics to determine
the appropriate, grammatical construction. Results therefore provided strong support
for the role of verb semantics in the avoidance of overgeneralisation errors.
An alternative, but related, semantic approach is not to organise verbs into
groups, but rather to take into account the consistency of the semantics of each verb
with the semantics of the construction into which it is being placed. In the current
chapter, we explore the influence of semantic verb-construction compatibility as a
continuum, in line with the approach taken by Ambridge et al. (e.g. 2011, 2014; see
also Bresnan, Cueni, Nikitina & Baayen, 2007). As verb compatibility with a
construction decreases, we predict (a) decreasing acceptability of ungrammatical,
overgeneralised sentences in our grammaticality judgment tasks, and (b) a decreasing
error rate in our production-priming study.
4.1.4 Recent evidence for the three accounts
Various studies have shown support for all three of the accounts investigated here,
with more recent studies tending to test more than one of these proposals in the same
experiment. Some of these papers have focussed on the transitive-
causative/intransitive-inchoative alternation (e.g. Braine, Brody, Fisch, Weisberger
& Blum, 1990; Brooks & Tomasello, 1999; Maratsos, Gudeman, Gerard-Ngo &
Dehart, 1987), although work has also been done on other construction pairs, such as
the locative (e.g. Ambridge, Pine & Rowland, 2012; Bidgood et al., 2014; Gropen et
al., 1991a, b) and dative alternations (e.g. Ambridge et al., 2014; Goldberg, 2011;
Stefanowitsch, 2008).
80
In a combined grammaticality judgment and production study, Boyd et al.
(2012) manipulated the frequency and construction type (intransitive only or
intransitive and periphrastic causative) in participants’ input for novel intransitive-
only verbs (e.g. The apple is yadding, The squirrel really made the apple yad). They
found evidence for independent effects of preemption and entrenchment. In support
of the preemption account, participants were less likely to produce and accept
transitive uses of the verbs when they heard them in both intransitive sentences and
the preempting periphrastic causative than when they had heard them in only the
intransitive (with the same total frequency). In support of the entrenchment account,
participants in the intransitive-only condition were less likely to produce transitive
sentences than a control group who were asked to produce sentences using the real
(alternating) verb bounce.
Ambridge et al. (2008; see also Ambridge et al., 2011) also used
grammaticality judgments with novel verbs, combined with known verbs. Support
for statistical learning accounts was found, as adults were less accepting of transitive
uses of high frequency than low frequency known intransitive-only verbs. For
example, *The man fell the cup (high frequency) was rated as less acceptable than
*The man tumbled the cup (low frequency). By manipulating semantics, the study
supported the predictions of the semantic verb class hypothesis. Participants were
taught novel verbs with semantics in line with one of Pinker’s (1989) intransitive-
only (intransitive-only or alternating in Ambridge et al., 2011) verb classes. For
example, animations showed the novel verb meek with semantics in line with
Pinker’s verbs of directed motion, such as fall and tumble. Participants were able to
use the semantic information to make grammaticality judgments for novel verbs in
line with those of semantically related known verbs. For example, The cup meeked
was judged as more acceptable than *The man meeked the cup.
The current study uses a regression design, following Ambridge, Pine and
Rowland’s (2012) grammaticality judgment study of the locative alternation (e.g.
ground locative Lisa sprayed the flowers with water/figure locative Lisa sprayed
water onto the flowers). While some verbs, like spray, are able to alternate between
the two locative constructions, others are only grammatical in the ground locative
(e.g. Lisa filled the cup with water vs. *Lisa filled water into the cup) or the figure
locative (e.g. Lisa poured water into the cup vs. *Lisa poured the cup with water).
Ambridge, Pine and Rowland presented participants with a series of sentences of
81
both types, containing ground-only, figure-only or alternating verbs. Adults
completed written questionnaires, whereas children (aged 5-6 and 9-10 years) heard
sentences, for a smaller set of verbs, accompanied by an animation, and gave their
judgments by placing a counter on a 5-point ‘smiley-face scale’.
The predictor variables used in Ambridge, Pine and Rowland (2012) were:
total verb frequency, calculated from a corpus, to test the entrenchment hypothesis;
verb frequency in the preempting construction (here, the grammatical locative
construction for non-alternating verbs), again calculated from a corpus, to test the
preemption hypothesis; and judgments from a separate group of adults about each
verb’s semantic properties, based on Pinker’s (1989) semantic verb class hypothesis,
to test this approach. The results were analysed using mixed effects linear regression
models. Regression allows each predictor variable to be entered as a continuous
variable, rather than the discrete variables of previous studies (e.g. high- vs. low-
frequency). It also allows the individual contribution of each variable to be assessed
in relation to that of the other variables. Using mixed effects models allows each of
the fixed effects (e.g. verb frequency, semantic score) to be analysed along with
random effects (e.g. participant), which may explain additional variation in the
results. Results showed significant effects of both overall verb frequency and
semantics, but frequency in the preempting construction had no dissociable effect
from overall verb frequency. Thus, support was found for the entrenchment and
semantic verb class hypotheses, but not preemption.
Following a similar methodology, Ambridge et al.’s (2014) investigation of
the dative construction found a significant effect of preemption, in all age groups
tested, in addition to the effects of entrenchment and semantics. This suggests that
the roles of entrenchment, preemption and semantic mechanisms may not play the
same role in the retreat from overgeneralisation errors for all construction types.
In summary, previous findings suggest roles for entrenchment, preemption
and semantics in the retreat from overgeneralisation errors resulting from the
transitive-causative/intransitive-inchoative alternation. However, the majority of the
studies reported here (a) use only a small number of verbs and/or (b) rely on
evidence from novel verbs. Both of these issues limit the generalisability of their
findings. Recent studies by Ambridge and colleagues (Ambridge, Pine & Rowland,
2012; Ambridge et al., 2014) have investigated the roles of all three mechanisms,
whilst overcoming these generalisability issues. Using grammaticality judgment data
82
in a regression design, these papers found distinct roles for semantics and
entrenchment, with preemption also playing a role in the retreat from dative, but not
locative, overgeneralisation errors. These findings provide a template for further
investigation of the transitive-causative/intransitive-inchoative alternation. However,
the grammaticality judgment methodology used by these may be of limited use, since
overgeneralisation is a phenomenon of production.
4.1.5 The current study
In this paper, we adopted a multi-method approach to the problem of the retreat from
overgeneralisation in the transitive-causative/intransitive-inchoative alternation,
using both grammaticality judgment and production tasks. We first sought to extend
the findings of Ambridge et al. (2012, 2014), from the locative and dative
constructions to the causative construction. To do this, in Experiment 1, we tested
adults using a grammaticality judgment task with a large number of verbs. In
Experiment 2, we used a scaled-down version of this task with both children and
adults to investigate how the relative contributions of the three mechanisms to
grammaticality ratings may change over the course of development. Having
established the factors that influence grammaticality judgments from both adults and
children, in Experiment 3, we used a production-priming methodology to investigate
whether these findings hold for children’s production. Specifically, we attempted to
elicit overgeneralisation errors from 5- to 6-year-olds and examine whether the
mechanisms under investigation (entrenchment, preemption and semantics) predicted
errors in the same way that they predicted grammaticality judgments. If this is the
case, further support will be added to the arguments presented in previous papers
using grammaticality judgment data only.
4.2 Methods
The current study consists of three experiments. Experiment 1 was a grammaticality
judgment study in which adults rated transitive and intransitive sentences containing
each of 180 verbs, of which 60 were transitive-only, 60 intransitive-only and 60
alternating. For all non-alternating verbs, a difference score was calculated by
subtracting each participant’s rating for the ungrammatical sentence from their rating
83
for its grammatical equivalent. The bigger this difference, the less acceptable
participants found the ungrammatical sentence to be, in comparison with its
grammatical counterpart. Experiment 2 repeated the grammaticality judgment
paradigm but with three age groups (adults, and children aged 5-6 and 9-10 years)
and a reduced set of verbs (40 of each type), as well as using animations to ensure
the meaning of each sentence was clear to the children. Experiment 3 used a
production-priming methodology with 5- to 6-year-olds to elicit overgeneralisations
with both transitive-only and intransitive-only verbs (e.g. *The ball kicked, *Homer
swam the fish). The same set of verbs was used as in Experiment 2, with the
alternating verbs used as fillers.
Before describing the individual experiments, we first outline the methods
used to create the predictor variables to test the entrenchment, preemption and
semantics hypotheses.
4.2.1 Frequency counts
In order to test the predictions of the entrenchment and preemption hypotheses, verb
frequency counts were taken from the British National Corpus (2007). The
entrenchment hypothesis (e.g. Braine & Brooks, 1995) posits an inference-from
absence mechanism to explain children’s retreat from overgeneralisation errors: the
more a verb is heard in the input, without being heard in the ungrammatical
construction, the stronger the inference that that verb-construction pairing must not
be possible. Thus, the entrenchment hypothesis predicts that the more a verb has
been heard regardless of the construction, the less acceptable it will be in
ungrammatical sentences, and the less likely children will be to produce an error with
that verb. Therefore, to test this account, we used counts of total verb frequency in
the corpus.
The preemption hypothesis (e.g. Goldberg, 1995), while related to the
entrenchment hypothesis, adds a semantic element: the more a verb is heard in
constructions with a roughly equivalent meaning to the ungrammatical construction,
the stronger the inference that the ungrammatical verb-construction pairing must not
be possible. Thus, the preemption hypothesis predicts that the more a verb is heard in
a competing construction with similar meaning, the less acceptable it will be in
ungrammatical sentences, and the less likely children will be to produce this type of
84
error with that verb. To test this account, for the transitive-only verbs, we used
counts of verb frequency in the passive (both full and truncated passives were
counted), as this is the construction proposed by Brooks and Tomasello (1999) as the
preempting construction for intransitivisation errors with transitive-only verbs (e.g.
*The ball kicked). Like the intransitive construction, the passive construction puts the
discourse focus on the patient by placing it first in the sentence (e.g. The plate broke;
The plate was broken [by Homer]). The truncated passive also allows the sentence to
exclude the agent altogether, as in the intransitive. In our corpus data, the majority of
passive sentences were truncated (92.15% of all passive uses of transitive-only
verbs), and thus the passive uses of these verbs are almost identical to the intransitive
construction, except for the auxiliary be. Again following Brooks and Tomasello
(1999), for intransitive-only verbs, we used counts of verb frequency in the
periphrastic causative (e.g. The man made the girl laugh), since this construction
expresses a similar meaning to the transitive-causative (e.g., *The man laughed the
girl) and overtly expresses both agent and patient. All verb frequency counts were
log transformed.
4.2.2 Semantic ratings
Under the semantics hypothesis, verb semantics determine the permissible
constructions for a particular verb, including the transitive and intransitive. Verbs
with similar semantics tend to behave similarly in terms of the constructions in which
they can appear. Pinker (1989) posited discrete classes of verbs, and identified the
key semantic features of each class. In line with more recent work (e.g. Ambridge,
Pine & Rowland, 2012; Ambridge et al., 2014), we treated verb semantics as a
continuum and created a measure of verb semantics by conducting a rating task to
determine the key characteristics of the transitive and intransitive constructions, as
well as the individual verbs that can, and cannot, appear in these constructions.
85
4.2.2.1 Method
4.2.2.1.1 Participants
The participants were 10 adults aged 20-25, all undergraduate students at the
University of Liverpool. They were each paid £50 for their participation. All
participants were monolingual speakers of English, and had no known language
impairments. They did not take part in the other experiments reported in this paper.
4.2.2.1.2 Test items
60 transitive-only verbs, 60 intransitive-only verbs and 60 verbs that can alternate
between the two structures were chosen as test items, based on Pinker (1989) and
Levin (1993). Table 4.1 lists the verbs used in the Experiments in this chapter.
86
Verb type
Also used in
Experiments
2 and 3 Verbs
Transitive-
only
Yes cut, slice, chop, mash, hit, strike, bite, peck, touch,
stroke, slash, saw, crush, squash, kick, tap, whack,
punch, nudge, kiss, kill, destroy, demolish, take,
bring, raise, hoist, lift, lower, leave, drop, tickle,
amuse, feed, delight, give, madden, lend, pay, offer
No slice, slay, murder, assassinate, slaughter, execute,
obliterate, poison, abandon, desert, gladden,
sadden, sicken, donate, sell, nauseate, bribe,
convince, enthuse, thrill, refund
Intransitive-
only
Yes go, come, rise, fall, tumble, ascend, descend, exit,
enter, arrive, eat, jump, hop, run, drink, talk, swim,
climb, sing, sleep, smile, cry ,laugh, frown, giggle,
chortle, chuckle ,grin, groan, moan, glow ,glitter,
leak, appear, disappear, vanish, die, stay, wait, live
No glisten, sweat, bleed, sparkle, twinkle, ooze, gush,
decease, perish, expire, emerge, materialize, lapse,
exist, cost, last, weigh, linger, sound, measure
Alternating Yes break, rip, shatter, smash, grow, change, bake, boil,
cook, fry, burn, split, tear, melt, crack, improve,
inflate, alter, shrink, freeze, crash, fold, crease,
deflate, defrost, dissolve, enlarge, expand, open,
close, snap, bend, slide, move, roll, bounce, turn,
begin, start, stop
No transform, divide, explode, fill, flood, stretch, heal,
evaporate, spin, rotate, float, hang, skid, cease,
commence, continue, end, finish, proceed, resume
Table 4.1. Verbs used in the semantic ratings task and Experiment 1. Those used in
Experiments 2 and 3 are also indicated.
4.2.2.1.3 Procedure
Each participant entered ratings on a spreadsheet. Both the verbs and the semantic
criteria were randomised separately for each participant. The instructions were as
follows:
On the following sheet is a list of 180 verbs. Each verb denotes an event
where one person or thing (A) causes another person or thing (B) to move or
change. In what follows, we always use A to denote the “causer” and B to
denote the thing that moves or changes. The spreadsheet also has a list of 26
different meanings, each of which applies to some verbs but not others. Your
87
task is to rate the extent to which each meaning applies to each verb, on a
scale of 1-9.
The procedure was illustrated with verbs and features not used in the task proper.
The 26 ‘meanings’ consisted of statements based on aspects of semantics thought to
be related to the transitive-causative and/or intransitive-inchoative constructions (see
Pinker, 1989; Levin, 1993). For example, Another person or thing comes into
physical contact with B is related to the transitive-causative construction; B is
something that can move by itself is related to the intransitive-inchoative
construction. Table 4.2 for a full list of these semantic criteria. At no point were the
verbs presented in sentences, nor were participants instructed to imagine them in
sentences. Mean ratings were calculated for each semantic criterion for each verb.
88
Another person or thing comes into physical contact with B
B allows or enables the movement/change (as opposed to actually initiating it)
B causes its own movement/change
B comes into (or goes out of) existence
B emits light/a sound/a substance
B expresses a particular emotion
B initiates the movement/change (as opposed to merely allowing or enabling it)
B is a human or animal
B is something that can move by itself
B moves in a particular direction
B moves in a particular manner
B moves to/from a particular location
B must be willing for the movement/change to occur
B remains in the same place whilst its parts move
B requires an external cause for the movement/change to happen
B undergoes a change-of state (e.g., solid to liquid; whole to pieces)
For the movement/change to be possible, B must be a particular type of thing
For the movement/change to be possible, B must have some particular property
The movement/change can occur even if B resists
The movement/change is internally caused (i.e., caused by B)
The movement/change is something that B does (as opposed to something that
"happens to" B)
The movement/change is voluntary as opposed to accidental on the part of B
The movement/change occurs to a particular degree/in a particular amount
The movement/change is something that "happens to" B (as opposed to something
that B does)
The verb denotes a particular property that B possesses (rather than
movement/change of B)
The verb denotes B being in a particular state (rather than movement/change of
B)
Table 4.2. Semantic criteria used in the semantic rating task.
89
Principal Components Analysis (PCA) with oblique rotation was used to
produce a number of composite factors, as many of the individual semantic criteria
may be highly correlated. This analysis was conducted in RStudio (version 0.97.551;
R version 3.1.1, R Core Team, 2014) following the procedure suggested by Field,
Miles and Field (2012, pp. 772-793) and using the following packages and functions:
corpcor (version 1.6.6; Schäfer, Opgen-Rhein, Zuber, Ahdesmäki, Duarte Silva &
Strimmer, 2013), GPArotation (version 2012.3-1; Bernaards & Jennrich, 2005),
psych (version 1.4.6; Revelle, 2014), kmo() (Kerns, 2007). The KMO statistic for the
data is 0.76 (‘good’ according to Field et al.) with all values for individual items
above the threshold of 0.5. Bartlett’s test of sphericity (χ2[df=190]=1970.70, p<0.001)
indicated that items were sufficiently correlated. Five components were retained,
based on parallel analysis (Horn, 1965). Details of how each of the original semantic
features load on each of the five composite semantic components are provided in
Table 4.3. Any loading of above 0.384 (or below -0.384) is statistically significant
(Stevens, 2012, p. 332). In subsequent analyses, all five of the extracted components
are entered into the models as predictor variables. Table 4.4 shows summary
information for each component.
90
Sp
ecific
pro
perty
/state
An
imate/
volitio
nal
patien
t
Extern
al
agen
t/cau
se
Man
ner/lo
catio
n
of m
ovem
ent
Deg
ree o
f
chan
ge
Another person or thing comes into physical
contact with B
-0.6 -0.01 0.76 -0.02 0.07
The movement / change is voluntary as
opposed to accidental on the part of B
-0.6 0.61 0.03 0 0.11
B must be willing for the movement / change
to occur
-0.2 0.72 -0.3 -0.1 0.09
B moves to/from a particular location -0.1 0.06 -0.2 0.85 -0.2
The movement / change is internally caused
(i.e., caused by B)
0.02 0.83 -0.1 0.14 -0.1
B requires an external cause for the
movement / change to happen
0.03 -0.4 0.73 0.02 0.09
B is something that can move by itself 0.04 0.88 -0.03 0.15 -0.04
B moves in a particular manner 0.09 0.11 0.18 0.88 0.25
B remains in the same place whilst its parts
move
0.12 0.23 0.39 -0.02 0.44
B comes into (or goes out of) existence 0.14 0.05 0.16 0.03 -0.9
B expresses a particular emotion 0.24 0.5 0.05 -0.4 0.16
B undergoes a change-of state (e.g., solid to
liquid; whole to pieces)
0.25 -0.3 0.62 0.21 -0.2
The movement / change occurs to a particular
degree/in a particular amount
0.28 -0.01 0.25 0.18 0.74
The movement / change can occur even if B
resists
0.3 -0.1 0.58 -0.2 -0.2
B emits light/a sound/a substance 0.39 0.18 0.33 -0.2 0.22
For the movement / change to be possible, B
must be a particular type of thing
0.43 0.56 0.21 -0.3 0
B allows or enables the movement / change
(as opposed to actually initiating it)
0.67 -0.2 -0.2 0.07 0.11
The verb denotes B being in a particular state
(rather than movement / change of B)
0.74 -0.1 0.11 0.02 -0.3
The verb denotes a particular property that B
possesses (rather than movement / change of
B)
0.74 -0.1 -0.2 -0.2 0.24
For the movement / change to be possible, B
must have some particular property
0.8 0.21 0.16 0.03 0.03
Table 4.3. Semantic factor loadings.
91
Component
label
Variance
explained
Verb type
related to
(Pinker, 1989)
Direction of
error-rate
prediction:
Transitive
verbs
Direction of
error-rate
prediction:
Intransitive
verbs
Specific
property/state
18% Intransitive + -
Animate/
volitional
patient
17% Intransitive + -
External cause/
agent
13% Transitive - +
Manner/
location of
movement
10% Intransitive + -
Degree of
change
10% Alternating unclear unclear
Table 4.4. Summary of composite semantic factors produced by PCA: component
label (given by us); the percentage of variance it explains; whether it is related to
transitive-only, intransitive-only or alternating verbs; and the direction in which it is
expected to predict the rates of intransitivisation errors with transitive-only verbs
(e.g. *The ball kicked) and transitivisation errors with intransitive-only verbs (e.g.
*Homer swam the fish). Specifically, “Direction of error rate prediction” refers to
whether we would expect more (+) or fewer (-) errors the higher a verb’s rating is on
each of the semantic components, in line with Pinker (1989).
Verbs with large values for specific property/state are those for which the
patient must have a specific property (e.g. being able to glow; glow) or be in a
particular state (e.g. solid; melt). Verbs with large values for animate/volitional
patient, are those for which the patient is animate (e.g. giggle) and/or able to do
things of its own accord (e.g. run). Verbs with large values for external cause/agent
are those which cannot take place without an external cause (e.g. heat; defrost) or
agent (e.g. an assassin; assassinate). Verbs with large values for manner/location of
movement are those which specify the manner of a movement (e.g. run) or the
location/direction of movement (e.g. rise). Verbs with large values for degree of
change are those for which the patient changes in some way (e.g. bend), but for
which that change does not involve coming into/going out of existence (verbs such as
die and materialise have large negative values).
92
4.2.3 Experiment 1 (adults – grammaticality judgments)
From previous studies looking at the dative and locative alternations, we know that
both corpus-derived entrenchment/preemption measures and experimentally derived
semantic measures can predict grammaticality ratings (e.g. Ambridge, Pine and
Rowland, 2012). Before examining children’s error production, we wanted to ensure
that these measures are also appropriate for investigating the transitive-
causative/intransitive-inchoative alternation. Therefore, in Experiment 1, adults rated
both transitive-causative and intransitive-inchoative sentences containing transitive-
only, intransitive-only or alternating verbs for grammaticality.
4.2.3.1 Method
4.2.3.1.1 Participants
The participants were 44 adults aged 20-25, all undergraduate students at the
University of Liverpool who had not taken part in the semantic ratings task. They
each were paid £10 for their participation. All participants were monolingual
speakers of English and had no known language impairments.
4.2.3.1.2 Test items
The same verbs were used as in the semantic ratings task. For each verb, transitive-
causative and intransitive-inchoative sentences were created as follows:
The man/woman/boy/girl [VERBed] the [object/person/animal] [modifying
phrase]
The [object/person/animal] [VERBed] [modifying phrase]
So, for example, the sentences for the alternating verb grow were The man grew the
flowers in the greenhouse and The flowers grew in the greenhouse.
93
4.2.3.1.3 Procedure
Participants were asked to judge the grammatical acceptability of the test sentences
described above. The sentences were presented, in a different random order for each
participant, on a spreadsheet. Each sentence was presented with a semantic ‘context’
in which the sentences were to be judged (in lieu of the animations used in the
subsequent studies), such as:
Context: MAN CAUSE GIRL SING ALL NIGHT
Sentence: The man sang the girl all night [or The girl sang all night]
Grammaticality ratings were given on a scale of 1-7, where 1 was completely
unacceptable and 7 was completely acceptable. Example dative sentences with low,
medium and high acceptability ratings (based on Ambridge, Pine, Rowland &
Chang, 2012) were given to illustrate how grammaticality ratings should be
distributed. Participants entered their rating for each sentence directly onto the
spreadsheet. From these raw ratings, a difference score was calculated for the
transitive-only and intransitive-only verbs by subtracting the rating for the
ungrammatical sentence from the rating for the grammatical sentence (transitive
minus intransitive for transitive-only verbs, intransitive minus transitive for
intransitive-only verbs). Alternating verbs were not included in the analyses as, by
definition, they are grammatical in both transitive and intransitive sentences. The
larger the difference score, the less acceptable the ungrammatical sentence, in
comparison with its grammatical counterpart. Previous studies (e.g. Ambridge et al.
2008) show that using a difference score as opposed to raw scores is important in
order to avoid a possible confound resulting from participants simply disliking
particular verbs, characters, events etc.
4.2.3.2 Results
Data were analysed separately for intransitivisation errors with transitive-only verbs
(e.g. *The ball kicked) and transitivisation errors with intransitive-only verbs (e.g.
*Homer swam the fish), allowing us investigate any possible differences in the
contribution of statistical and semantic mechanisms to adults’ judgments of errors of
94
each type. Analyses were carried out using linear mixed effects regression models in
RStudio (version 0.97.551; R version 3.1.1, R Core Team, 2014) with the lmer
function of the lme4 package (version 1.1-7, Bates, Maechler, Bolker & Walker,
2014). Predictor variables were total verb frequency (the entrenchment measure),
verb frequency in the preempting construction (passive for transitive-only verbs,
periphrastic causatives for intransitive-only verbs) and all five semantic components
extracted from the PCA (described above). Random intercepts for verb and
participant were included in the model, with by-participant random slopes for as
many predictor variables as possible while still allowing the models to converge
(Barr, Levy, Scheepers & Tily, 2013). The predictor variables were initially
standardised (transformed into z scores) before being entered into the model. Model
summaries are shown in Table 4.5.
95
Variable Transitive-
only verbs
Intransitive-
only verbs
Fixed effects Intercept 3.00 (0.16) 3.26 (0.20)
Estimate (SE) Entrenchment 0.67 (0.22) 0.45 (0.14)
Preemption -0.45 (0.23) 0.09 (0.14)
Specific property/state -0.39 (0.12) 0.11 (0.14)
Animate/volitional patient 0.21 (0.12) 0.37 (0.14)
External agent/cause 0.21 (0.11) -0.02 (0.13)
Manner/location of movement 0.06 (0.12) -0.31 (0.14)
Degree of change -0.14 (0.11) 0.10 (0.13)
Random effects Verb (intercept) 0.49 (0.70) 0.62 (0.79)
Variance (SD) Participant (intercept) 0.61 (0.78) 1.26 (1.12)
Participant – Entrenchment n/a 0.04 (0.19)
Participant – Preemption 0.04 (0.19) n/a
Participant – Specific
property/state
0.04 (0.21) 0.03 (0.17)
Participant –
Animate/volitional patient
n/a 0.06 (0.25)
Participant – External
agent/cause
0.05 (0.22) 0.02 (0.16)
Participant – Manner/location
of movement
0.02 (0.16) n/a
Participant – Degree of change 0.02 (0.14) 0.07 (0.27)
Model summary AIC 10924 10986
BIC 11106 11168
Log Likelihood -5430.8 -5461.8
Deviance 10862 10942
Table 4.5. Model summary: Experiment 1, all verbs.
Model comparisons were computed using the anova function to determine
significance levels for each of the predictor variables. Each variable was removed in
reverse order, so the semantic component Degree of change was removed first and
the Entrenchment measure last. Entrenchment is first in the model (and removed last
in the model comparison process) as it is currently the mechanism with the most
supporting evidence in the literature and, therefore, was deemed to be the most
widely-accepted in the field. The other predictor variables would need to
demonstrate that they could operate over and above entrenchment. Preemption was
entered next, as it is also well-defined in the literature and the frequency counts are a
subset of those used for the entrenchment predictor. Finally, the semantic predictors
were entered in the order of the amount of variance they explained in the PCA (see
Table 4.4).
96
The results of the model comparison analysis for transitive-only verbs are
shown in Table 4.6 and for intransitive-only verbs in Table 4.7. As Table 4.6 shows,
for intransitivisation errors with transitive-only verbs (e.g. *The ball kicked),
significant predictor variables were total verb frequency, verb frequency in the
preempting passive construction, and the semantic components specific
property/state and external agent/cause. Both of these semantic components predict
results in the expected direction: specific property/state, a feature of intransitive
verbs, negatively predicts difference scores for intransitivisation errors with
transitive-only verbs, whereas external agent/cause, a feature of transitive verbs,
positively predicts these scores. These findings therefore provide support for the
entrenchment and semantic hypotheses, although preemption was a negative
predictor – the opposite of our expectation. We return to this issue in the Discussion.
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 24 10953 11094 -5452.6 10905
Entrenchment 25 10940 11087 -5444.9 10890 15.46 1 8.4e-5
***
Preemption 26 10935 11088 -5441.6 10883 6.57 1 0.010*
Specific
property/state
27 10927 11086 -5436.5 10873 10.15 1 0.0014
**
Animate/
volitional
patient
28 10926 11091 -5435.2 10870 2.63 1 0.10
External
agent/cause
29 10922 11092 -5431.9 10864 6.55 1 0.011*
Manner/
location of
movement
30 10924 11100 -5431.8 10864 0.25 1 0.62
Degree of
change
31 10924 11106 -5430.8 10862 1.97 1 0.16
Table 4.6. Model comparisons: Experiment 1, transitive-only verbs.
97
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 24 11009 11150 -5480.7 10961
Entrenchment 25 10997 11144 -5473.7 10947 13.96 1 1.9e-4
***
Preemption 26 10996 11149 -5472.0 10944 3.32 1 0.069
Specific
property/state
27 10996 11155 -5471.0 10942 2.12 1 0.15
Animate/
volitional
patient
28 10987 11152 -5465.5 10931 10.94 1 9.4e-4
***
External agent/
cause
29 10989 11159 -5465.5 10931 0.053 1 0.82
Manner/
location of
movement
30 10984 11161 -5462.1 10942 6.67 1 0.0098
***
Degree of
change
31 10986 11168 -5461.8 10942 0.78 1 0.38
Table 4.7. Model comparisons: Experiment 1, intransitive-only verbs.
As Table 4.7 shows, for transitivisation errors with intransitive-only verbs
(e.g. *Homer swam the fish), significant predictor variables were total verb
frequency, and semantic components animate/volitional patient and manner/location
of movement. Both of these semantic components predict results in the expected
direction: animate/volitional patient, a feature of intransitive verbs, positively
predicts difference scores for transitivisation errors with intransitive-only verbs,
whereas manner/location of movement, a feature of alternating verbs, negatively
predicts these scores. Again, these results provide support for both the entrenchment
and semantics hypotheses. Verb frequency in the periphrastic was only a marginally
significant predictor (p=0.07).
In summary, the results from Experiment 1 show mixed support for the three
theories under investigation. While there is strong support for both the entrenchment
and semantics hypotheses, no support was found for the preemption hypothesis.
4.2.4 Experiment 2 (adults and children – grammaticality judgments)
Experiment 1 showed that total verb frequency (entrenchment) and verb semantics
significantly predict difference scores for ungrammatical uses of both transitive-only
and intransitive-only verbs, for adults in a grammaticality judgment task, with mixed
98
findings for the preemption hypothesis. However, our main goals in this paper were
to explore the influence of these three mechanisms on overgeneralisation errors (a)
developmentally and (b) in production. Thus before turning to production, we ran a
second judgment study (Experiment 2) with both adults and children. This study had
two aims. The first was to investigate the respective influences of entrenchment,
preemption and semantics across development (which requires a judgment task, as
older children and adults are unlikely to produce overgeneralisation errors). The
second was to obtain judgment data that could be subsequently compared to
production data obtained from children of the same age. This is important for
generalising the findings from previous child judgment studies to the main
phenomenon of interest: children’s overgeneralisation errors in production.
4.2.4.1 Method
4.2.4.1.1 Participants
The participants were 96 children aged 5-6 (5;3-6;5, M=5;10), 96 children aged 9-10
(9;4-10;6, M=9;11), and 24 adults aged 18-25. The children were recruited from
primary schools in the North West of England. The adults were all undergraduate
students at the University of Liverpool and received course credit for their
participation. They had not taken part in Experiment 1 or the semantic ratings task.
All participants were monolingual speakers of English, and had no known language
impairments.
4.2.4.1.2 Test items
The number of verbs was reduced from 180 in Experiment 1 (60 each of transitive-
only, intransitive-only and alternating) to 120 (40 of each type). The verbs used were
a subset of those previously used, chosen to be those most likely to be known to
young children. The number of test items was reduced to ensure the task was not too
long for the attention span of the younger group of children. The sentences used were
similar to those in Experiment 1, with the exception that, in order to make the study
more child-friendly, the generic terms describing the agents in transitive sentences
(e.g. the girl) were replaced with names of familiar cartoon characters, such as Lisa
dropped the ball to the floor.
99
Animations were created, using Anime Studio Pro 5.5, for each of the actions.
The transitive and intransitive sentences for each verb were recorded by a native
speaker of British English. The audio file played automatically after the animated
action ended. Identical animations were used for the transitive and intransitive
versions of each sentence. The use of animations ensured that the veracity of the
sentences would not be in doubt and that participants’ judgments would therefore be
based on the grammaticality of the sentences, something that we have previously
found to be important when testing young children (e.g. Ambridge et al., 2008).
4.2.4.1.3 Procedure
Test sentences and their accompanying animations were presented to participants
using VLC Media Player. Grammaticality judgments were given on a 5-point
‘smiley-face’ judgment scale (see e.g., Ambridge et al., 2008), shown in Figure 4.1.
Adults watched the full set of animations, in a pseudo-random order such that no two
sentences containing the same verb were presented consecutively, in small groups of
up to 10 participants. Adults marked their responses (individually) on an answer
sheet containing one smiley-face scale for each sentence. Due to constraints on
attention span, children were tested individually on one quarter of the sentences (60
in total), split over two days. Thus, 24 children rated each verb. Each child was tested
on transitive and intransitive versions of sentences containing 10 each of transitive-
only, intransitive-only and alternating verbs. Sentences were again presented in
pseudo-random order. Children gave their responses by placing a green or red
counter (indicating broadly grammatical or broadly ungrammatical, respectively)
onto a single, larger smiley-face scale (see Figure 4.1). The experimenter noted down
responses by hand. As in Experiment 1, a difference score was calculated for the
transitive-only and intransitive-only verbs by subtracting the rating for the
ungrammatical sentence from the rating for the grammatical sentence, resulting in a
difference score for the ungrammatical sentence relative to its grammatical
counterpart.
100
Figure 4.1. 5-point ‘smiley-face’. The face on left represents a completely
ungrammatical sentence, the face on the right a completely grammatical sentence and
the remaining faces a rating between these two extremes.
4.2.4.2 Results
Results were analysed using the same method as in Experiment 1. Predictor variables
in the initial analysis were age group, total verb frequency (the entrenchment
measure), verb frequency in the periphrastic causative/passive (the preemption
measure) and all five semantic components extracted from the PCA, as well as
interaction terms for age group by each of the other predictor variables. As in
Experiment 1, random intercepts were included for verb and participant, with by-
participant random slopes for as many predictor variables as possible while still
allowing the model to converge. Interaction terms, however, were not included as
random slopes as the model could not converge when this was the case. A summary
of the model is shown in Table 4.8. The results of model comparisons are shown in
Table 4.9 for transitive-only verbs and Table 4.10 for intransitive-only verbs.
101
Variable Transitive-
only verbs
Intransitive-
only verbs
Fixed effects Intercept 1.59 (0.10) 1.99 (0.10)
Estimate (SE) Age group 0.07 (0.07) 0.97 (0.08)
Entrenchment 0.20 (0.19) 0.55 (0.15)
Preemption -0.01 (0.18) -0.15 (0.13)
Specific property/state -0.16 (0.10) -0.14 (0.14)
Animate/volitional patient 0.31 (0.10) 0.04 (0.13)
External agent/cause 0.05 (0.10) 0.21 (0.12)
Manner/location of movement -0.12 (0.10) -0.19 (0.12)
Degree of change 0.05 (0.13) 0.17 (0.13)
Age group:Entrenchment 0.15 (0.08) -0.09 (0.06)
Age group:Preemption -0.01 (0.08) 0.04 (0.04)
Age group:Specific
property/state
0.07 (0.04) 0.08 (0.06)
Age group:Animate/ volitional
patient
0.13 (0.05) -0.02 (0.06)
Age group:External
agent/cause
0.01 (0.05) 0.06 (0.05)
Age group:Manner/ location of
movement
-0.08 (0.04) -0.15 (0.04)
Age group:Degree of change 0.12 (0.06) -0.10 (0.05)
Random effects Verb (intercept) 0.24 (0.49) 0.28 (0.53)
Variance (SD) Participant (intercept) 0.37 (0.61) 0.40 (0.64)
Participant – Age group n/a 0.07 (0.27)
Participant – Entrenchment 0.08 (0.28) 0.09 (0.30)
Participant – Preemption 0.06 (0.24) n/a
Participant – Specific
property/state
n/a 0.07 (0.26)
Participant –
Animate/volitional patient
0.06 (0.24) 0.05 (.023)
Participant – External
agent/cause
0.04 (0.21) 0.11 (0.34)
Participant – Manner/location
of movement
0.02 (0.13) n/a
Participant – Degree of change 0.05 (0.21) 0.02 (0.14)
Model summary AIC 10380 10477
BIC 10654 10751
Log Likelihood -5143.7 -5192.5
Deviance 10288 10385
Table 4.8. Model summary: Experiment 2, all verbs and age groups, with by-age
interactions.
102
Predictor df AIC BIC logLik deviance χ2 χ
df
p
value
Intercept 31 10463 10648 -5200.6 10401
Age group 32 10398 10589 -5167.2 10334 66.77 1 3.0e-16
***
Entrenchment 33 10397 10594 -5165.6 10331 3.19 1 0.07
Preemption 34 10398 10600 -5165.0 10330 1.20 1 0.27
Specific
property/state
35 10400 10608 5164.8 10380 0.41 1 0.52
Animate/
volitional
patient
36 10390 10604 -5158.9 10318 11.70 1 6.3e-4
***
External
agent/cause
37 10391 10612 -5158.6 10317 0.61 1 0.43
Manner/
location of
movement
38 10391 10618 -5157.6 10315 2.05 1 0.15
Degree of
change
39 10393 10626 -5157.5 10315 0.11 1 0.74
Age group:
Entrenchment
40 10394 10632 -5157.0 10314 1.16 1 0.28
Age group:
Preemption
41 10394 10638 -5155.9 10312 2.18 1 0.14
Age group:
Specific
property/state
42 10393 10643 -5154.3 10309 3.08 1 0.08
Age group:
Animate/
volitional
patient
43 10384 10640 -5148.7 10298 11.17 1 8.3e-4
***
Age group:
External
agent/cause
44 10385 10648 -5148.7 10297 0.08 1 0.77
Age group:
Manner/
location of
movement
45 10382 10650 -5145.7 10292 5.93 1 0.015
*
Age group:
Degree of
change
46 10380 10654 -5143.7 10288 3.99 1 0.046
*
Table 4.9. Model summary: Experiment 2, transitive-only verbs, all age groups, with
by-age interactions.
103
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 31 10588 10773 -5263.0 10526
Age group 32 10516 10706 -5225.9 10452 74.39 1 2.2e-16
***
Entrenchment 33 10508 10705 -5221.0 10442 9.64 1 0.0019
***
Preemption 34 10510 10713 -5221.0 10442 0.032 1 0.86
Specific
property/state
35 10512 10720 -5220.8 10442 0.38 1 0.54
Animate/
volitional
patient
36 10509 10724 -5218.7 10437 4.28 1 0.039 *
External
agent/cause
37 10508 10728 -5216.8 10434 3.85 1 0.0498
*
Manner/
location of
movement
38 10506 10733 -5215.1 10430 3.34 1 0.068
Degree of
change
39 10507 10740 -5214.6 10429 1.01 1 0.32
Age group:
Entrenchment
40 10501 10740 -5210.7 10421 7.76 1 0.0054
**
Age group:
Preemption
41 10502 10746 -5209.8 10420 1.75 1 0.19
Age group:
Specific
property/state
42 10485 10736 -5200.6 10401 18.42 1 1.8e-5
***
Age group:
Animate/
volitional
patient
43 10488 10744 -5200.9 10402 0 1 1
Age group:
External
agent/cause
44 10488 10751 -5200.2 10400 1.30 1 0.25
Age group:
Manner/
location of
movement
45 10478 10747 -5194.2 10388 12.11 1 5.0e-4
***
Age group:
Degree of
change
46 10477 10751 -5192.5 10385 3.26 1 0.071
Table 4.10. Model summary: Experiment 2, intransitive-only verbs, all age groups,
with by-age interactions.
For intransitivisation errors with transitive-only verbs (e.g. *The ball kicked),
model comparisons revealed main effects of age group and the semantic component
animate/volitional patient, and significant interactions for age by three of the
104
semantic components (animate/volitional patient, manner/location of movement and
degree of change). Neither the entrenchment nor the preemption predictors yielded
significant main effects or interactions, although the main effect for entrenchment
was marginal (p=0.07). For transitivisation errors with intransitive-only verbs (e.g.
*Homer swam the fish), model comparisons revealed main effects of age, total verb
frequency (entrenchment) and two of the semantic components (animate/volitional
patient and external agent/cause). Significant interactions for age by total verb
frequency and age by two of the semantic components (specific property/state and
manner/location of movement) were also found.
The significant interactions indicate that different age groups were using
semantic information differently when making grammaticality judgments for
intransitivisation errors with transitive-only verbs (e.g. *The ball kicked), and both
verb frequency and semantic information differently for transitivisation errors with
intransitive-only verbs (e.g. *Homer swam the fish). We therefore carried out
additional analyses by age group to investigate these differences further.
4.2.4.2.1 5- to 6-year-olds
Table 4.11 shows the model summaries for both verb types. Tables 4.12 and 4.13
show the results of the model comparisons for transitive-only and intransitive-only
verbs, respectively. For transitive-only verbs, none of the variables significantly
predicted difference scores, although the semantic component specific property/state
was marginally significant in the expected direction. For intransitive-only verbs, total
verb frequency (entrenchment) and the semantic component animate/volitional
patient significantly predicted difference scores in the expected direction.
105
Variable Transitive-
only verbs
Intransitive-
only verbs
Fixed effects Intercept 0.83 (0.11) 0.95 (0.13)
Estimate (SE) Entrenchment 0.12 (0.19) 0.63 (0.19)
Preemption -0.10 (0.18) -0.23 (0.15)
Specific property/state -0.19 (0.10) -0.25 (0.18)
Animate/volitional patient 0.09 (0.09) 0.14 (0.17)
External agent/cause 0.07 (0.10) 0.15 (0.15)
Manner/location of movement -0.02 (0.10) -0.05 (0.15)
Degree of change -0.10 (0.14) 0.28 (0.17)
Random effects Verb (intercept) 0.12 (0.34) 0.34 (0.58)
Variance (SD) Participant (intercept) 0.33 (0.57) 0.35 (0.59)
Participant – Entrenchment 0.27 (0.52) 0.19 (0.44)
Participant – Preemption 0.13 (0.36) n/a
Participant – Specific
property/state
n/a 0.12 (0.34)
Participant –
Animate/volitional patient
n/a 0.08 (0.29)
Participant – External
agent/cause
0.03 (0.18) 0.19 (0.44)
Participant – Manner/location
of movement
0.07 (0.26) n/a
Participant – Degree of change 0.14 (0.38) 0.11 (0.33)
Model summary AIC 3739.3 3947.3
BIC 3889.5 4098.8
Log Likelihood -1838.7 -1942.7
Deviance 3677.3 3885.3
Table 4.11. Model summary: Experiment 2, 5- to 6-year-olds, all verbs.
Predictor df AIC BIC logLik deviance χ2 χ df p value
Intercept 24 3733.0 3849.3 -1842.5 3685.0
Entrenchment 25 3734.0 3855.2 -1842.0 3684.0 0.94 1 0.33
Preemption 26 3735.5 3861.5 -1841.8 3683.5 0.49 1 0.48
Specific
property/state
27 3734.5 3865.3 -1840.2 3680.5 3.08 1 0.08
Animate/
volitional
patient
28 3735.1 3870.8 -1839.5 3679.1 1.37 1 0.24
External
agent/cause
29 3735.9 3876.4 -1839.0 3677.9 1.18 1 0.28
Manner/
location of
movement
30 3737.9 3883.9 -1839.0 3667.9 0.01 1 0.93
Degree of
change
31 3739.3 3889.5 -1838.7 3677.3 0.57 1 0.45
Table 4.12. Model comparisons: Experiment 2, 5- to 6-year-olds, transitive-only
verbs.
106
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 24 3995.6 4072.1 -1953.8 3907.6
Entrenchment 25 3949.3 4070.7 -1949.7 3899.3 8.29 1 0.0040
**
Preemption 26 3950.7 4076.9 -1949.4 3898.7 0.60 1 0.44
Specific
property/state
27 3951.3 4082.3 -1948.6 3897.3 1.48 1 0.22
Animate/
volitional
patient
28 3946.6 4082.5 -1945.3 2890.6 6.67 1 0.0098
**
External
agent/cause
29 3946.5 4087.3 -1944.3 3888.5 2.06 1 0.15
Manner/
location of
movement
30 3948.5 4094.1 -1944.2 3888.5 0.057 1 0.81
Degree of
change
31 3947.3 4097.8 -1942.7 3885.3 3.13 1 0.077
Table 4.13. Model comparisons: Experiment 2, 5- to 6-year-olds, intransitive-only
verbs.
4.2.4.2.2 9- to 10-year-olds
Table 4.14 shows the model summaries and Tables 4.15 and 4.16 show the results of
the model comparisons for transitive-only and intransitive-only verbs, respectively.
For transitive-only verbs, total verb frequency (entrenchment) and the semantic
component animate/volitional patient significantly predicted difference scores, while
verb frequency in the preempting passive construction was not significant. However,
the semantic prediction is not in the expected direction: animate/volitional patient is
a factor related to intransitivity (Pinker, 1989) and, therefore, was expected to be a
negative predictor of difference scores for intransitive uses of transitive-only verbs.
(The more a transitive verb’s semantics are related to having an animate or volitional
patient, the more highly that verb was expected to be rated in the intransitive. The
difference between judgment scores for the [‘grammatical’] transitive and
[‘ungrammatical’] intransitive sentences was therefore expected to be smaller for
verbs rated more highly for this semantic component. However, this was not the case
here: the higher a verb’s rating on the semantic component, the larger the difference
score was.)
107
Variable Transitive-
only verbs
Intransitive-
only verbs
Fixed effects Intercept 1.65 (0.16) 2.11 (0.14)
Estimate (SE) Entrenchment 0.22 (0.30) 0.58 (0.19)
Preemption 0.06 (0.29) -0.11 (0.16)
Specific property/state -0.20 (0.17) -0.15 (0.19)
Animate/volitional patient 0.47 (0.16) -0.06 (0.18)
External agent/cause 0.01 (0.16) 0.24 (0.15)
Manner/location of movement -0.10 (0.16) -0.17 (0.15)
Degree of change 0.14 (0.21) 0.18 (0.17)
Random effects Verb (intercept) 0.60 (0.78) 0.42 (0.65)
Variance (SD) Participant (intercept) 0.42 (0.64) 0.49 (0.70)
Participant – Entrenchment 0.02 (0.14) 0.08 (0.28)
Participant – Preemption 0.09 (0.31) 0.11 (0.34)
Participant – Specific
property/state
n/a 0.08 (0.29)
Participant –
Animate/volitional patient
0.08 (0.28) 0.11 (0.34)
Participant – External
agent/cause
0.07 (0.26) 0.09 (0.29)
Participant – Manner/location
of movement
n/a n/a
Participant – Degree of change 0.06 (0.24) 0.14 (0.37)
Model summary AIC 3591.3 3536.9
BIC 3742.2 3721.9
Log Likelihood -1764.7 -1730.5
Deviance 3529.3 3460.9
Table 4.14. Model summary: Experiment 2, 9- to 10-year-olds, all verbs.
Predictor df AIC BIC logLik deviance χ2 χ
df
p
value
Intercept 24 3596.5 3713.3 -1774.3 3548.5
Entrenchment 25 3594.6 3716.3 -1772.3 3544.6 3.94 1 0.047 *
Preemption 26 3595.1 3721.6 -1771.5 3543.1 1.51 1 0.22
Specific
property/state
27 3596.8 3728.2 -1771.4 2542.8 0.25 1 0.62
Animate/
volitional
patient
28 3586.6 3722.9 -1765.3 3530.6 12.21 1 4.8e-4
***
External
agent/cause
29 3588.6 3729.8 -1765.3 3530.6 0.02 1 0.89
Manner/
location of
movement
30 6589.8 3735.8 -1764.9 3529.8 0.78 1 0.38
Degree of
change
31 3591.3 3742.2 -1764.7 3529.3 0.49 1 0.48
Table 4.15. Model comparisons: Experiment 2, 9- to 10-year-olds, transitive-only
verbs.
108
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 31 3540.6 3691.5 -1739.3 3478.6
Entrenchment 32 3533.9 3689.7 -1735.0 3469.9 8.65 1 0.0033
**
Preemption 33 3535.9 3696.5 -1735.0 3469.9 0.0036 1 0.95
Specific
property/state
34 3537.2 3702.7 -1734.6 3469.2 0.70 1 0.40
Animate/
volitional
patient
35 3537.7 3708.0 -1733.8 3467.7 1.54 1 0.21
External
agent/cause
36 3535.6 3710.8 -1731.8 3463.6 4.09 1 0.043 *
Manner/
location of
movement
37 3536.2 3716.3 -1731.1 3462.2 1.41 1 0.23
Degree of
change
38 3536.9 3721.9 -1730.5 3460.9 1.26 1 0.26
Table 4.16. Model comparisons: Experiment 2, 9- to 10-year-olds, intransitive-only
verbs.
For intransitive-only verbs, total verb frequency and the semantic component
external agent/cause significantly predicted difference scores, while verb frequency
in the preempting periphrastic causative construction was not significant. Again, the
semantic component external agent/cause predicted results in the opposite direction
to what was expected: the more a verb’s semantics were in line with this factor,
which relates to transitivity, the greater participants’ difference scores for
ungrammatical transitive uses of these verbs. Taken together with the findings for
transitive-only verbs, it seems that 9- to 10-year-olds are not using semantic
information as predicted, although this information does seem to affect their
judgments in some way.
4.2.4.2.3 Adults
Table 4.17 shows the model summaries and Tables 4.18 and 4.19 show the results of
the model comparisons for transitive-only and intransitive-only verbs, respectively.
For transitive-only verbs, the semantic components animate/volitional patient and
manner/location of movement were both significant predictors of difference scores,
while total verb frequency (entrenchment) and verb frequency in the preempting
109
passive construction were marginally significant predictors. The semantic component
animate/volitional patient relates to intransitivity but here is a positive predictor of
difference scores for ungrammatical intransitive uses of transitive-only verbs, the
opposite direction to our original expectation.
Variable Transitive-
only verbs
Intransitive-
only verbs
Fixed effects Intercept 2.18 (0.13) 2.81 (0.13)
Estimate (SE) Entrenchment 0.30 (0.19) 0.43 (0.15)
Preemption -0.03 (0.18) -0.12 (0.13)
Specific property/state -0.05 (0.11) -0.07 (0.15)
Animate/volitional patient 0.29 (0.10) 0.04 (0.14)
External agent/cause 0.12 (0.10) 0.22 (0.12)
Manner/location of movement -0.19 (0.10) -0.35 (0.12)
Degree of change 0.13 (0.13) 0.07 (0.13)
Random effects Verb (intercept) 0.23 (0.48) 0.29 (0.45)
Variance (SD) Participant (intercept) 0.19 (0.43) 0.18 (0.42)
Participant – Entrenchment 0.06 (0.24) 0.02 (0.13)
Participant – Preemption 0.03 (0.17) n/a
Participant – Specific
property/state
0.02 (0.14) n/a
Participant –
Animate/volitional patient
n/a 0.01 (0.08)
Participant – External
agent/cause
0.01 (0.08) 0.00 (0.06)
Participant – Manner/location
of movement
0.02 (0.12) 0.01 (0.09)
Participant – Degree of change n/a n/a
Model summary AIC 2874.4 2689.2
BIC 3025.3 2810.9
Log Likelihood -1406.2 -1319.6
Deviance 2812.4 2689.2
Table 4.17. Model summary: Experiment 2, adults, all verbs.
110
Predictor df AIC BIC logLik deviance χ2 χ
df
p
value
Intercept 24 2887.0 3003.8 -1419.5 2839.0
Entrenchment 25 2886.0 3007.7 -1418.0 2836.0 3.01 1 0.08
Preemption 26 2885.1 3011.6 -1416.5 2833.1 2.91 1 0.09
Specific
property/state
27 2887.1 3018.5 -1416.5 2833.1 0.01 1 0.91
Animate/
volitional
patient
28 2876.1 3012.4 -1410.0 2820.1 12.98 1 3.1e-4
***
External
agent/cause
29 2876.7 3017.8 -1409.3 2818.7 1.45 1 0.23
Manner/
location of
movement
30 2873.5 3019.6 -1406.8 2813.5 5.11 1 0.024 *
Degree of
change
31 2874.4 3025.3 -1406.2 2812.4 1.14 1 0.29
Table 4.18. Model comparisons: Experiment 2, adults, transitive-only verbs.
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 18 2701.8 2789.3 -1332.9 2665.8
Entrenchment 19 2700.4 2792.9 -1331.2 2662.4 3.32 1 0.068
Preemption 20 2702.2 2799.6 -1331.1 2662.2 0.18 1 0.67
Specific
property/state
21 2699.8 2802.0 -1328.9 2657.8 4.47 1 0.034 *
Animate/
volitional
patient
22 2967.4 2804.4 -1326.7 2653.4 4.41 1 0.036 *
External
agent/cause
23 2694.2 2806.1 -1324.1 2648.2 5.18 1 0.023*
Manner/
location of
movement
24 2687.6 2804.4 -1319.8 2639.6 8.57 1 0.0034
**
Degree of
change
25 2689.2 2810.9 -1319.6 2689.2 0.40 1 0.53
Table 4.19. Model comparisons: Experiment 2, adults, intransitive-only verbs.
For intransitive-only verbs, results show that semantic components specific
property/state, animate/volitional patient, external agent/cause and manner/location
of movement all significantly predicted difference scores, while total verb frequency
(entrenchment) again did so only marginally, and preemption not at all. As with the
transitive-only verbs, adults’ difference scores were not always predicted in the
expected direction by the semantic components. Predictions in the opposite direction
to that expected were found with the components specific property/state, which
111
relates to intransitivity and was therefore expected to be a positive predictor, and
external agent/cause, which related to transitivity and was therefore expected to be a
negative predictor. Like the 9- to 10-year-olds, then, and contrary to findings in
Experiment 1, adults in this study do not seem to be using semantic information as
expected.
4.2.4.3 Discussion
Overall, the results from Experiment 2 show some support for the entrenchment
hypothesis, but none for the preemption hypothesis. The entrenchment measure (total
verb frequency) significantly predicted difference scores for transitive-only verbs for
the 9- to 10-year-olds and marginally for adults. For intransitive-only verbs,
entrenchment was a significant predictor for both groups of children, although not for
adults (unlike Experiment 1). Therefore, it is likely that the influence of
entrenchment on grammaticality judgments decreases with age, explaining the
interaction of age by entrenchment for intransitive-only verbs in the original model
in Table 4.8. In contrast, the preemption predictors (verb frequency in the passive for
transitive-only verbs and in the periphrastic causative for intransitive-only verbs) did
not significantly predict any results (although preemption was a marginally
significant predictor for adults), contrary to findings with studies investigating other
constructions.
Strong support for the importance of semantics was suggested in Experiment
2. For all age groups, at least one of the semantic measures significantly predicted
difference scores for ungrammatical transitive uses of intransitive-only verbs (e.g.
*Homer swam the fish) and, for the two older age groups, for ungrammatical
intransitive uses of transitive-only verbs (e.g. *The ball kicked). However, the older
children and adults did not always seem to be using semantic information as
expected, with predictions for some of the semantic components going in the
opposite direction to our expectations. Age by semantics interactions in the original
model (Table 4.8) appear to be due to different semantic components playing
different roles over the different age groups, with semantics in general apparently
becoming more important as age increases (and as reliance on frequency information
appears to decrease).
112
In addition, the finding that none of the variables was able to predict
grammaticality judgments for intransitivisation errors with transitive-only verbs in
the youngest age group indicates that these children might be unable to interpret
errors of this type, therefore rendering children incapable of making systematic
judgments on the acceptability of these sentences.
4.2.5 Experiment 3 (children – production-priming)
Experiments 1 and 2 confirmed that adults and older children use a combination of
verb frequency and semantic information when making grammaticality judgments of
intransitivisation and transitivisation overgeneralisations. For 5- to 6-year-olds,
however, the results so far have suggested that both entrenchment and semantics, but
not preemption, influence judgments for transitive overgeneralisations with
intransitive-only verbs, with no significant findings for intransitive
overgeneralisations with transitive-only verbs. While these results tell us something
about children’s metalinguistic knowledge of such errors, Experiment 3 allows us to
investigate our main question of interest: whether children will actually produce
overgeneralisation errors with transitive-only and intransitive-only verbs and, if so,
what factors predict by-verb differences in error rates: entrenchment, preemption or
semantics. Argument structure overgeneralisation errors in naturalistic diary studies
are sparse and limited to those that the transcriber happens to notice. So, to test our
three hypotheses, we used an elicited production-priming task aimed at encouraging
the production of overgeneralisation errors in 5- to 6-year-old children.
4.2.5.1 Method
4.2.5.1.1 Participants
The participants were 64 children aged 5-6 (5;2-6;4, M=5;8) recruited from primary
schools in the North West of England. All were monolingual speakers of English and
had no known language impairments. None of these children had participated in
Experiment 2.
113
4.2.5.1.2 Test items
Test items were the same as Experiment 2, with the addition of a single alternating
verb (float, produced by the experimenter only), added for the purpose of the bingo
game described below.
4.2.5.1.3 Materials
The 120 verbs (40 each of transitive-only, intransitive-only and alternating) were
split into four sets, each containing 20 alternating verbs and 10 each of the transitive-
only and intransitive-only verbs. Alternating verbs were therefore used twice as
many times in total as fixed-transitivity verbs, since they were used in both priming
conditions. Each child received a single verb set for their target verbs. The first
experimenter used 10 transitive-only and 10 intransitive-only verbs for the prime
sentences; there was therefore no overlap between the verbs received by the child
and those produced by the experimenter.
4.2.5.1.4 Procedure
The aim of this experiment was to encourage children to produce both
intransitivisation errors with transitive-only verbs (e.g. *The ball kicked; cf. Homer
kicked the ball) and transitivisation errors with intransitive-only verbs (e.g. *Homer
swam the fish; cf. The fish swam). In order to do this, we used a production-priming
methodology in which an experimenter produced grammatical intransitive-inchoative
sentences to encourage the child to use this construction with transitive-only verbs.
On a separate day, children were primed with grammatical transitive-causative
sentences to encourage the production of transitivisation errors with intransitive-only
verbs. Examples of trials in each prime condition are given below, with the target
error we were attempting to elicit.
114
(1) Intransitive prime condition (transitive-only target verbs)
a. Experimenter 2 (clue words): lightbulb, glow
b. Experimenter 1: The lightbulb glowed
c. Experimenter 2 (clue words): ball, hit
d. Child: *The ball hit
(2) Transitive prime condition (intransitive-only target verbs)
a. Experimenter 2 (clue words): lift, bag
b. Experimenter 1: Bart lifted the bag
c. Experimenter 2 (clue words): wait, boy
d. Child: *Lisa waited the boy
Each child participated on two occasions, on separate days. In each session,
children took turns with an experimenter to describe a series of animations. These
animations were presented using Processing (www.processing.org). Both
experimenter and child were given ‘clue words’ by a second experimenter to
encourage them to use the intended verb. The clue words consisted of the verb
followed by the direct object, when transitive sentences were being primed, or the
subject followed by the verb, when intransitive sentences were being primed. The
second experimenter noted down children’s responses, although all sessions were
also audio-recorded using Audacity in order to check responses later if there was any
doubt about what the child had said.
Half of the children received transitive primes on the first day and intransitive
primes on the second, and vice versa for the other children. The first three pairs of
animations were training trials containing only transitive-only or intransitive-only
verbs for both experimenter and child, whichever the child was to be primed with on
that day. These verbs were not in the child’s verb set, nor were they used as primes
by the experimenter in that child’s test trials. Twenty test trials then followed, with
the experimenter continuing to use transitive-only or intransitive-only verbs,
depending on prime condition. The experimenter produced only grammatical
sentences. In contrast, half of the target verbs given to the children were alternating
verbs (and would therefore be grammatical whether the child produced a transitive or
an intransitive sentence) and half were intransitive-only or transitive-only, whichever
was the opposite of the prime condition. For these trials, if the child produced a
115
sentence using the same construction as that with which they had been primed, an
overgeneralisation error would result.
In order to motivate the children to produce the sentences, a ‘bingo game’
was used (as in Rowland, Chang, Ambridge, Pine & Lieven, 2012). Each time
Experimenter 1 or the child produced a sentence, Experimenter 2 (who could not see
the computer screen) looked for a matching bingo card. In fact, Experimenter 2 had
all of the bingo cards and whether or not the card was given to Experimenter 1/the
child was predetermined: the games were fixed so that the child always won both
games on the first day, lost the first game on the second day (in order to maintain
tension) and the won the final game. This manipulation required an extra trial for
Experimenter 1 only, on Day 2, always with the (alternating) verb float. Each ‘game’
lasted for ten trials each, in order to keep the child’s attention and motivation.
4.2.5.2 Results
Children’s responses were coded for sentence type: transitive (active), intransitive,
passive (full or truncated), periphrastic causative, other use of the verb, irrelevant
(target verb not included/no response). As we are investigating overgeneralisation
errors, the constructions of interest were intransitive uses of transitive-only verbs and
transitive uses of intransitive-only verbs. Sentences were only included in the
analysis if the child used the target verb in his/her response, with error rate calculated
as a proportion of errors from the total number of responses that included the target
verb. Replacement of NPs with pronouns or generic terms was allowed (e.g. the dad
hit the ball for Homer hit the ball; it fell for the cup fell), as were changes in
tense/aspect (e.g. Homer hit/hits/was hitting the ball), morphological
overgeneralisations (e.g. The ball hitted) and additional modifying phrases (e.g. He
kicked the ball in the goal). The mean number of sentences of each type produced by
each child is shown in Figure 4.2, of a possible maximum of 10.
116
Figure 4.2. Mean number of sentences of each type produced by each child (N=64),
with 95% CIs. Total number of sentences produced per child per verb type is 10.
Black bars represent verbs that are considered grammatical only in the transitive
construction (transitive-only verbs) and grey bars represent verbs that are considered
grammatical only in the intransitive construction (intransitive-only verbs). The
constructions given on X-axis are those produced by the child.
The binary dependent variable for this experiment is the child’s response:
overgeneralisation error (1) or other use of the target verb (0), with all responses in
which the child did not use the target verb excluded from the analysis. As this is a
binary variable, results were analysed using the glmer function of the lme4 package
(version 1.1-7, Bates et al., 2014), with family=binomial. Predictor variables were
the same as Experiment 1. Random intercepts for verb and participant were included
in the models, although all random slopes were removed in order to enable the
models to converge.
As can be seen in Figure 4.2, our production-priming method succeeded in
eliciting a large number of overgeneralisation errors, in both directions, from
children age 5-6 years. The success of this task gives us a unique insight into the
0
1
2
3
4
5
6
7
8
9
10M
ean
res
pon
ses
by c
hil
d
Sentence type
Transitive-only
Intransitive-only
117
mechanisms that influence children’s error production for a large number of children
with a large number of known verbs, a big advantage over corpus or diary studies, or
elicitation tasks using a small set of novel verbs. The large number of errors children
produced enabled us to examine by-verb differences, as well as testing if the
mechanisms under investigation (entrenchment, preemption and semantics) pattern
similarly in their predictions of error rates to their predictions of grammaticality
judgments.
Table 4.20 shows the model summaries and Tables 4.21 and 4.22 show the
results of the model comparisons for transitive-only and intransitive-only verbs,
respectively. For intransitivisation errors with transitive-only verbs (e.g.*The ball
kicked), semantic component external agent/cause was the only significant predictor
of error rate. This prediction is in the expected direction. Neither the entrenchment
measure nor the preemption measure predicted error rates.
Variable Transitive-
only verbs
Intransitive-
only verbs
Fixed effects Intercept 1.24 (0.49) 1.98 (0.52)
Estimate (SE) Entrenchment -0.51 (0.49) -1.00 (0.32)
Preemption -0.04 (0.48) -0.24 (0.27)
Specific property/state 0.20 (0.24) 0.20 (0.26)
Animate/volitional patient -0.18 (0.26) -0.16 (0.27)
External agent/cause -0.78 (0.28) -0.15 (0.24)
Manner/location of
movement
0.01 (0.27) 0.24 (0.23)
Degree of change -0.16 (0.29) -0.72 (0.27)
Random effects
Variance (SD)
Verb (intercept) 0.85 (0.92) 0.12 (0.35)
Participant (intercept) 8.86 (2.98) 8.65 (2.94)
Model summary AIC 461.09 391.07
BIC 502.56 432.25
Log Likelihood -220.55 -185.53
Deviance 441.09 371.07
Table 4.20. Model summary: Experiment 3, all verbs.
118
Predictor df AIC BIC logLik deviance χ2 χ
df
p
value
Intercept 3 457.93 470.37 -225.97 451.93
Entrenchment 4 459.39 475.97 -225.69 451.39 0.55 1 0.46
Preemption 5 461.38 482.11 -225.69 451.38 0.00 1 0.95
Specific
property/state
6 461.68 486.56 -224.84 449.68 1.70 1 0.19
Animate/
volitional
patient
7 463.36 492.39 -224.68 449.36 0.32 1 0.57
External
agent/cause
8 457.44 490.61 -220.72 441.44 7.29 1 0.0049
**
Manner/
location of
movement
9 459.40 496.71 -220.70 441.40 0.04 1 0.83
Degree of
change
10 461.09 502.56 -220.55 441.09 0.30 1 0.58
Table 4.21. Model comparisons: Experiment 3, transitive-only verbs.
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 3 412.77 425.12 -203.38 406.77
Entrenchment 4 402.03 418.50 -197.01 394.03 12.74 1 3.6e-4
***
Preemption 5 398.45 419.04 -194.22 388.45 5.58 1 0.018*
Specific
property/state
6 400.45 425.16 -194.22 388.45 0.00 1 0.97
Animate/
volitional
patient
7 395.34 424.16 190.67 381.34 7.11 1 0.0077
**
External
agent/cause
8 395.43 428.38 -189.72 379.43 1.90 1 0.17
Manner/
location of
movement
9 396.74 433.81 -189.37 378.74 0.69 1 0.41
Degree of
change
10 391.07 432.25 -185.53 371.07 7.68 1 0.0056
**
Table 4.22. Model comparisons: Experiment 3, intransitive-only verbs.
For transitivisation errors with intransitive-only verbs (e.g. *Homer swam the
fish), total verb frequency (entrenchment), verb frequency in the periphrastic
causative (preemption) and the semantic components animate/volitional patient and
degree of change all significantly predicted children’s error rates in production. The
semantic components animate/volitional patient and degree of change both relate to
intransitivity and are negative predictors of error rates: the greater the extent to which
119
the semantics of an intransitive verb are in line with these components, the less likely
children are to overgeneralise that verb into the transitive-causative construction.
That is, children make fewer transitivisation errors with intransitive-only verbs the
higher a verb is rated on these semantic components, usually associated with
intransitive-only verbs (e.g. Pinker, 1989). This indicated that, the more typically
intransitive a verb is, the less likely children are to overgeneralise it into the
transitive construction (e.g. *Homer swam the fish).
In summary, entrenchment and preemption measures predict error rates for
transitivisation errors with intransitive-only verbs, but not the converse error. In
terms of semantics, for both verb types, overgeneralisation errors were significantly
predicted, in the expected direction, by individual semantic components. The results
for this production study pattern in a similar way to the grammaticality judgment task
in Experiment 2, thus validating the use of both methodologies here, as well as in
previous studies.
4.3 General discussion
This paper used a multi-methodological approach to investigate three hypotheses that
aim to explain children’s retreat from overgeneralisation errors. The entrenchment
hypothesis posits that children retreat from or avoid error through encountering verbs
which are used multiple times in grammatical constructions (e.g. Bart kicked the ball,
The fish swam). Children then infer from their absence in the input that
ungrammatical verb-construction combinations (e.g. *The ball kicked, *Homer swam
the fish) are not possible, and the more they hear the grammatical sentences, the
greater the strength of this inference. The preemption hypothesis posits that hearing
the verb used only in grammatical constructions with a similar meaning (e.g. The
ball was kicked, Homer made the fish swim) will help children retreat from or avoid
error, as they have a viable alternative available to express the same meaning. Again,
the prediction is that the more children hear a given verb in these preempting
constructions, the better able they will be to retreat from (or avoid) error. Finally, the
semantics hypothesis predicts that, as children’s knowledge of verb semantics and
how these fit with the semantics of the construction increases, their errors will
decrease because they will become aware that certain verb-construction pairings are
not semantically compatible (e.g. it is not possible to say *The ball kicked because an
120
external agent is required; it is not possible to say *Homer swam the fish because the
swimming motion is internally caused).
Previous studies (e.g. Ambridge et al., 2008; 2014; Bidgood et al., 2014;
Brooks & Tomasello, 1999) have shown support for a mixture of the three
hypotheses under investigation, with a number of different constructions. However,
most of these have used only a small number of verbs and/or relied on the use of
novel verbs. Studies using larger numbers of verbs have employed a grammaticality
judgment methodology, whereas overgeneralisation is really a production
phenomenon. The three experiments presented here investigated intransitivisation
errors with transitive-only verbs (e.g. *The ball kicked) and transitivisation errors
with intransitive-only verbs (e.g. *Homer swam the fish), using grammaticality
judgment tasks with adults and children (Experiments 1 and 2) and a production-
priming task with 5- to 6-year-olds (Experiment 3). Grammaticality judgments allow
for a comparison with findings from previous studies with other constructions (e.g.
Ambridge, Pine & Rowland, 2012), with the priming task giving us a crucial insight
into children’s production of errors. Overall, the results provided strong support for
entrenchment and semantic mechanisms, suggesting both are likely to play a role in
the retreat from overgeneralisation for errors involving the transitive-
causative/intransitive-inchoative alternation. Evidence for a preemption mechanism
is found, but only in the production task.
Echoing the findings of previous papers (e.g. Ambridge et al., 2008; Boyd et
al., 2012; Naigles et al., 1992), clear support was found for a role played by
entrenchment in the grammaticality judgments of adults and older children with both
intransitivisation errors of transitive verbs (e.g. *The ball kicked) and transitivisation
errors of intransitive verbs (e.g. *Homer swam the fish). For the younger group of
children, the operation of this mechanism is clear with transitivisation errors of
intransitive verbs in the judgment task and, importantly, in the production task.
Therefore, the entrenchment mechanism clearly has a role to play in the retreat from
overgeneralisation for errors involving the transitive-causative/intransitive-inchoative
alternation.
In contrast to previous findings (e.g. Boyd et al., 2012; Brooks & Tomasello,
1999), the present study found only marginal support for the preemption hypothesis.
In Experiment 1, verb frequency in the preempting construction predicted errors in
an unexpected direction for intransitivisation errors with transitive-only verbs, and
121
only marginally predicted transitivisation errors with intransitive-only verbs.
Preemption was unable to predict difference scores in the grammaticality judgment
task for any age group in Experiment 2. However, in the production study,
preemption did explain additional variance over and above that explained by
entrenchment for transitivisation errors with intransitive-only verbs. This is likely to
be due to a task effect (see also Blything, Ambridge & Lieven, 2014). In production,
all possible constructions are competing to express the message the child wants to
convey, thus forcing a choice between the alternative constructions – transitive vs.
periphrastic in this case. In contrast, judgment tasks do not require participants to
consider the alternative ways in which the message could have been conveyed. Other
possibilities relate to the frequency of the preempting constructions: very low
numbers of uses of the verbs in these constructions were found in the corpus. It could
be that: (1) the verbs we chose for this study did not have a wide enough spread of
uses in the preempting constructions for the mechanism’s operation to be observed;
(2) the corpus we used was too small to obtain accurate counts of preempting
constructions; or, most likely, (3) preemption is genuinely unable to operate for the
overgeneralisation errors under investigation due to the low frequency of the
preempting constructions in the input. In any case, while the preemption mechanism
is not well supported in our grammaticality judgment tasks, importantly, it does play
a role in children’s error production and, therefore, their retreat from error.
As in several previous studies (e.g. Ambridge et al., 2008; Brooks &
Tomasello, 1999), evidence for a semantics mechanism is strong in all three
experiments presented here, with semantic components predicting difference scores
in both grammaticality judgment tasks and, importantly, in the production task.
However, the direction of prediction was not always that which was expected for
adults and older children in Experiment 2. Since results were as expected in
Experiment 1, the smaller number of verbs in Experiment 2 could have caused
problems with the analysis, possibly related to the fact that the PCA from which the
semantic components were drawn involved the entire set of verbs used in Experiment
1. Another explanation lies in the process of labelling of the semantic components.
While conducting PCA is essential in order to (a) make the number of components
manageable, and (b) collapse individual features explaining overlapping variance, the
labelling of components drawn from a PCA is inherently problematic due to the large
122
number of individual features loading onto each of these, thus making them difficult
to interpret.
To investigate this possibility further, Table 4.23 shows the direction of
prediction for each semantic component. The Table illustrates the fact that each
semantic component either consistently predicts difference scores positively (and
error rates negatively) or negatively (and error rates positively), irrespective of verb
type. So, while these factors sometimes predicted results in the opposite direction to
expectations, the direction of predictions are consistent across experiments and age
groups. This could indicate that, rather than creating semantic components that
indicate how ‘transitive’ or ‘intransitive’ a verb is, we have a set of semantic
components that predict whether or not a verb can alternate between the two
constructions under investigation. Regardless of whether this explanation is correct,
the fact remains that verb semantics are able to predict both grammaticality
judgments and error rates for both types of overgeneralisation investigated here,
lending strong support to the semantics hypothesis. A mechanism involving verb
semantics therefore clearly has a role to play in the retreat from overgeneralisation
for errors involving the transitive-causative/intransitive-inchoative alternation.
Exp. 1
Exp. 2
(Age 5)
Exp. 2
(Age 9)
Exp. 2
(Adults)
Exp. 3
Semantic
component
TR
N
INT
TR
N
INT
TR
N
INT
TR
N
INT
TR
N
INT
Specific property/
state (INT)
- n/a n/a n/a n/a n/a n/a - n/a n/a
Animate/volitional
patient (INT)
n/a + n/a + + n/a + + n/a -
External agent/
cause (TRN)
+ n/a n/a n/a n/a + n/a + - n/a
Manner/location of
movement (ALT)
n/a - n/a n/a n/a n/a - - n/a n/a
Degree of change
(INT)
n/a n/a n/a n/a n/a n/a n/a n/a n/a -
Table 4.23. Direction of prediction of semantic components in all experiments.
Expected relationship with (in)transtitivity is indicated in the first column: INT =
intranstivity; TRN = transitivity; ALT = alternating.
123
4.3.1 Changes across development
The results of Experiment 2 suggest that the importance of entrenchment decreases
with development, as verb frequency effects are observed for both age groups of
children but not for adults. However, entrenchment does predict grammaticality
judgments for the adults tested in Experiment 1, in line with previous findings (e.g.
Ambridge et al., 2008; 2011). Experiment 1 included more verbs than Experiment 2,
with the majority of these additional verbs being of low frequency. This suggests the
existence of some sort of ceiling effect in the usefulness of statistical mechanisms:
the more times a verb has been encountered in the input, the less of an influence
hearing it yet more times will have on judgments of grammatical acceptability,
perhaps with the effect levelling off for the most frequent verbs. The children in
Experiment 2 may not have reached this threshold with the verbs included in that
study, whereas adults have not reached the threshold with the lower frequency verbs
in Experiment 1.
Results in Experiment 2 also suggest that, as the importance of entrenchment
decreases over time, the role of semantics comes to play a more important role. This
could imply simply that semantics explain a larger portion of variance once
frequency effects are no longer important. However, it is likely that semantics are
better-learnt in this oldest age group, therefore allowing the mechanism to operate
fully in adults when it was unable to do so in children. This suggestion is consistent
with Blything et al.’s (2014) investigation of un- prefixation errors (e.g. *unopen,
*unsqueeze).
4.3.2 How do children retreat from overgeneralisation errors?
The results of the three experiments reported in this paper have demonstrated strong
support for the entrenchment and semantics hypotheses in the retreat from
overgeneralisation errors. Both factors were significant predictors for grammaticality
judgment ratings in different age groups, and error rate in a production-priming task.
The preemption hypothesis also received some support, notably in the critical
production task.
These findings provide additional evidence for the claims of other recent
work that suggest the need for an account of language acquisition, including the
124
retreat from overgeneralisation errors, that integrates semantic and statistical
accounts (e.g. Ambridge et al., 2008; 2011). Taking evidence from the current paper
along with previous papers allows us to take a wider view of the mechanisms
involved in the retreat from overgeneralisation errors by looking across different
constructions. Preemption is important when the erroneous construction and its
preempting alternative have an almost identical meaning, as with the prepositional
object and double object datives (e.g. Homer gave the book to Marge vs. Homer gave
Marge the book; Ambridge et al., 2014) and especially the case of un- prefixation
errors (e.g. close preempts *unopen; Ambridge, 2013; Blything et al., 2014).
Preemption also works best when the preempting construction is relatively high
frequency, as in the dative (see above) and locative alternations (e.g. Lisa sprayed
the flowers with water vs. Lisa sprayed water onto the flowers; Ambridge et al.,
2011).
The current has demonstrated that the low frequencies of the passive and
periphrastic causative, and a potential lack of close semantic compatibility with the
constructions they could preempt, work against preemption in the case of the
transitive-causative/intransitive-inchoative alternation. Entrenchment, on the other
hand, has been shown to work well in the absence of a closely preempting alternative
construction of high frequency. Semantics plays an important role in the retreat from
overgeneralisation when there is a clear split between the two constructions involved
in the alternation – this is the case in the current paper, but less so for the alternative
dative constructions, for example. Finally, when children are forced to choose
between possible alternative constructions, as in the production-priming task used in
Experiment 3, but not in the grammaticality judgment task used in Experiment 2, the
role of preemption plays a more important role. This echoes findings in Blything et
al.’s (2014) paper on un- prefixation errors.
4.3.3 Conclusion
This paper tested the predictions of three accounts of the retreat from
overgeneralisation in the transitive-causative construction: entrenchment, preemption
and semantics. Using a combination of grammaticality judgment and production-
priming methods, testing a large number of verbs and treating predictor variables as
continuous has allowed for a particularly strong test of these hypotheses. Results
125
strongly support both the entrenchment and semantics hypotheses, and the
preemption hypothesis to a lesser degree. We therefore suggest that a successful
account of the retreat from overgeneralisation must therefore posit a role for both
statistical and semantic mechanisms, such as the FIT account (Ambridge & Lieven,
2011; Ambridge et al., 2011).
127
Chapter 5: Children’s and adults’ passive syntax is semantically constrained:
Evidence from syntactic priming
5.0 Fit within the thesis
Chapters 3 and 4 investigated the roles of semantics, entrenchment and preemption in
children’s retreat from argument structure overgeneralisation errors, finding support
for semantics and entrenchment in particular. However, the retreat from
overgeneralisation is a relatively narrow focus for any theory. Finding support for
these mechanisms in language acquisition more generally is therefore necessary if
they are to be taken seriously. This chapter focuses on one of these mechanisms –
semantics – investigating its role in children’s acquisition of the passive construction.
It presents a production-priming study with both adults and children, using the
priming method of Chapter 4. It used two approaches to semantics: the more
traditional semantic verb class approach taken Chapter 3, and the fine-grained
continuum approach to semantics taken in Chapter 4. The idea of a continuum of
semantic fit between the verb and the construction being a key factor in explaining
by-verb differences is therefore developed here. If this approach can successfully
explain differences in the acquisition of the passive construction, then semantic fit is
not just relevant to the retreat from overgeneralisation, but to the acquisition of
syntax more generally. The passive construction is particularly appropriate as a test
of our semantics approach (verb-in-construction compatibility) as previous studies
(e.g. Maratsos et al., 1985; Pinker et al., 1987) have found that children have
difficulties with certain verbs in the passive until a relatively advanced age.
This chapter is currently being prepared for submission to a peer-reviewed
journal. For the purposes of publication, it has been written as a response to
Messenger et al. (2012). (Note that, unlike the previous chapters, verb frequency
information has not been included in the analyses in the current chapter, as
Messenger et al. did not include frequency information in their analyses.
Nevertheless, see Ambridge, Bidgood, Pine, Rowland & Freudenthal, 2015, for a
similar analysis that included frequency information but, nevertheless, came to
similar conclusions regarding the role of semantics.) Messenger et al.’s conclusion
that children have an abstract representation of the passive by the age of 3 years is
supported by our findings, but the details of their conclusion are questioned, as our
128
modified method yielded by-verb differences that could not be seen given the
method used by Messenger et al. Although the initial focus of this chapter is on early
semantic restriction versus early abstraction accounts, its focus is on finding an
alternative approach to these, based on verb-in-construction compatibility. It
therefore challenges generativist-nativist assumptions surrounding early abstraction,
and illustrates how the approach developed throughout the Thesis so far can explain
more than simply children’s retreat from overgeneralisation errors.
5.1 Introduction
A key question in language acquisition is the extent to which children’s earliest
knowledge of syntax is abstract, with some researchers arguing that it is largely
abstract from the beginning of multi-word speech (e.g. Wexler, 1998; Gertner et al.,
2006) and others suggesting that this knowledge is initially lexically or semantically
restricted (e.g. Schlesinger, 1988; Tomasello, 2003). A construction that has often
been studied with regard to this debate is the English passive, which makes for a
particularly useful test case as it is one of few constructions for which children (and
even adults; e.g. Dąbrowska & Street, 2006) make errors in comprehension (e.g.
Maratsos et al., 1985; Sudhalter & Braine, 1985; Gordon & Chafetz, 1990; Fox &
Grodzinksky, 1998; Meints, 1999; Hirsch & Wexler, 2006).
Many studies of the passive have used syntactic priming (see Branigan, 2007,
and Pickering & Ferreira, 2008, for reviews). This method is particularly suitable for
investigating the abstractness of linguistic representations, as participants cannot be
primed to produce a syntactic structure for which they do not have an abstract
representation (Branigan et al., 1995). Priming has been used to investigate the
passive in both adults (e.g. Bock, 1986; Bock & Loebell, 1990; Bock, Loebell &
Morey, 1992) and children (e.g. Lempert, 1990; Savage, Lieven, Theakston &
Tomasello, 2003; 2006; Huttenlocher et al., 2004; Bencini & Valian, 2008;
Messenger, Branigan & McLean, 2011a; 2011b; Messenger et al., 2012). While a
number of previous studies have found that performance differs on a verb-by-verb
basis, supporting the idea of a semantic restriction on the passive, others have
observed excellent performance across verbs, supporting the early abstraction
approach. In the present article, we aim to reconcile these disparate findings by
means of a new passive priming study with children (aged 4-6) and adults.
129
5.1.1 The early semantic restriction account
Previous evidence for the early semantic restriction account comes from studies
showing that children appear to struggle more with comprehension and production of
passive sentences with mental state verbs (e.g. Bob was seen by Wendy) than with
actional verbs (e.g. Bob was hit by Wendy) (e.g. Maratsos et al., 1985; Pinker et al.,
1987). Maratsos et al. (1985) used a sentence comprehension task to compare young
children’s (aged from 4 to 11 years) understanding of active and passive sentences
using (a) prototypical actional verbs (agent-patient [AP] verbs, e.g. hold) and (b)
mental verbs (experiencer-theme [ET] verbs, e.g. like). All children performed above
chance with both verb types in active sentences (e.g. Batman holds Donald [Duck],
Batman sees Donald) and with AP verbs in passive sentences (e.g. Donald is held by
Batman). In contrast, children as old as 7 years struggled with ET verbs in passive
sentences (e.g. Donald is seen by Batman). The authors concluded that children
begin with a prototypical passive construction, which involves verbs with a high
degree of transitivity, such as AP verbs. As they get older, their representation of the
passive is broadened to include other verbs, before reaching the adult-like state in
which most (although not all) transitive verbs can be used in the passive (cf. e.g.
Donald was seen by Batman but not *1lb was weighed by the package). Additional
evidence that early knowledge of the passive is semantically restricted comes from
further studies of comprehension (e.g. Sudhalter & Braine, 1985; Gordon & Chaftez,
1990; Meints, 1999) and production (e.g. Tomasello, Brooks and Stern, 1998;
Meints, 1999), naturalistic data (e.g. Israel, Johnson & Brooks, 2000) and syntactic
priming (e.g. Savage et al., 2003).
An alternative interpretation of these data is that, although children’s
knowledge of syntax is, in general, abstract, the passive is subject to some additional
construction-specific difficulty. For example, Borer and Wexler’s explanation (1987)
derives from how passive sentences are thought to be ‘generated’ via a generative
grammar. As the object of an active sentence is ‘raised’ to become the subject of the
passive sentence, an A-chain (Argument chain) is formed. This causes the correct
form of the auxiliary verb to be to be inserted. Borer and Wexler (1987; see also
Hirsch & Wexler, 2006) is that young children are unable to form A-chains and that
their correct interpretation of truncated passives with some verbs (e.g. AP verbs)
reflects the fact that they interpret these as adjectival uses (e.g. Bob was hugged).
130
Fox and Grodzinsky (1998) claim that young children are able to form A-chains, and
show that they can correctly interpret truncated passives of all verb types, as well as
full passives with actional verbs. In contrast, the young children in their study were
unable to correctly interpret non-actional full passives (e.g. Bob was seen by Wendy).
Fox and Grodzinsky propose that this is because children are not yet able to assign
the correct thematic role to the NP in the by-phrase.
5.1.2 The early abstraction account
Previous evidence for the early abstraction account comes from syntactic priming
studies showing that, even for children as young as 3;0, hearing a passive sentence
increases the likelihood of subsequently producing a passive sentence with a
different verb (e.g. Savage et al., 2003; 2006; Huttenlocher et al., 2004; Bencini &
Valian, 2008; Messenger et al., 2011a). While these studies provide evidence that
young children’s knowledge of the passive is at least partly abstract, they do not
provide direct evidence against the early semantic restriction account, as none of
them directly compared children’s performance with verbs of different semantic
types.
Recently, Messenger et al. (2012) conducted production-priming and forced-
choice comprehension tasks designed to investigate this question. Specifically,
Messenger et al.’s aim was to investigate the possibility that children’s early
knowledge of the passive is indeed abstract, and that the by-verb differences
observed in previous studies (e.g. Maratsos et al., 1985) could be explained by task
effects. They argue that, because depicting non-actional events in experimental
materials is more difficult than depicting events involving AP verbs, poor picture
recognition may account for these results. Messenger et al.’s comprehension task
replicated the findings of previous studies, with young children showing more
accurate comprehension of passives with both agent-patient (AP; e.g. hug) and
theme-experiencer (TE; e.g. annoy) verbs (both of which involve ‘affectedness’ of
the passive subject) than with experiencer-theme (ET; e.g. see) verbs (in which the
passive subject may not be affected). In contrast to Maratsos et al.’s (1985) findings
(see also Hirsch & Wexler, 2006), Messenger et al. (2012) also found this pattern
with active sentences, supporting their claim that it is ET verbs (or depictions of ET
verbs) that are problematic for children, rather than ET passives.
131
To investigate adults’ and children’s abstract knowledge of the passive,
Messenger et al. (2012) conducted a production-priming study. Participants took
turns with an experimenter to describe pictures in which an animal was doing
something to a human (e.g. a sheep hitting/shocking a girl). The experimenter
produced an active or passive sentence containing either an AP or a TE verb
(Experiment 1) or a TE or an ET verb (Experiment 2). Participants’ pictures always
depicted an actional event, designed to be described using an AP verb (e.g. scratch,
wash, hug). Messenger et al. found no difference in the rate of passives produced
following different prime types, and this finding was the same for both adults and
children. They therefore concluded that children have adult-like abstract knowledge
of passive syntax by 3-4 years of age.
5.1.3 A third possibility: The semantic construction prototype account
A third possibility, not investigated in previous studies, is that, while even young
children have an abstract representation of the passive construction, this
representation is nevertheless semantically constrained. A recent study suggests that
adults’ knowledge of the passive, while undisputedly abstract, may take the form of a
semantic prototype construction. Ambridge, Bidgood, Pine, Rowland and
Freudenthal (2015) conducted graded grammaticality judgment and timed forced-
choice comprehension studies (both of which have the advantage of yielding a
continuous, rather than binary, dependent measure). After controlling for both overall
verb frequency and verb frequency in the passive construction, Ambridge et al. found
that the verbs’ affectedness ratings – designed to capture the semantics of the passive
construction – significantly predicted both adults’ judgments of grammatical
acceptability and their reaction times for choosing the correct animation in the
forced-choice comprehension task. (Note that almost all AP and TE verbs were rated
highly for affectedness. This contrasts with ET verbs, the vast majority of which had
low affectedness ratings.) Importantly, semantic effects for both grammaticality
judgments and reaction times were significantly larger for the passive than the active
construction. The implication is that the early semantic restriction account is right in
highlighting by-verb semantic differences in the passive (which persist into
adulthood), while the early abstraction account is right in highlighting the fact that,
nevertheless, both adults and children have an abstract verb-general representation of
132
the construction. This suggests the need for an approach that integrates aspects of
both the early semantic restriction account and the early abstraction account. One
such approach is the semantic construction prototype account (e.g. Ibbotson &
Tomasello, 2009): the passive construction is associated with the semantics of
affectedness of the passive subject (e.g. Pinker et al., 1987), and the greater the
extent to which a particular verb has compatible semantics, the greater the
acceptability and ease of comprehension of the relevant passive sentence. For
example, in the sentence Bob was scared by Wendy, Bob is definitely affected by the
event; thus scare is highly compatible with the semantics of the passive. In contrast,
in the sentence Bob was seen by Wendy, Bob may or may not be affected by the
event and may even be unaware that it has occurred; thus see is less compatible with
the semantics of the passive.
5.1.4 The present study
The aim of the current study is to test the predictions of (a) the early semantic
restriction account (b) the early abstraction account and (c) the semantic construction
prototype account. The study uses a production-priming task similar to that used by
Messenger et al. (2012), but with one crucial difference: Messenger et al. varied the
semantic type of the prime verb (AP/TE/ET) while holding constant the semantic
type of the target verb (AP). We vary the semantic type of the target verb
(AP/TE/ET), while holding constant the semantic type of the prime verb (AP). This
small manipulation results in a more sensitive test of by-verb differences. From other
studies, we know that the identity of the target verb can affect the size of the priming
effect (see Coyle & Kaschak, 2008; Peter, Chang, Pine, Blything & Rowland, 2015).
By using as targets only AP verbs – often suggested to be the prototypical verb type
for passive sentences and, therefore, presumably the easiest for children to produce –
Messenger et al. may have reduced the likelihood of observing by-verb differences.
That is, even a prime verb whose semantics are less than fully compatible with the
semantics of the passive construction (e.g. an ET verb) may still yield a “good
enough” passive to prime production of an “easy” AP passive. By reversing the
design of Messenger et al., the present study investigates whether, when primed with
a prototypical passive, the extent to which children are able to produce a passive
varies as a function of the semantics of the target verb (AP/TE/ET).
133
If the early abstraction account is correct, our results should essentially
replicate the findings of Messenger et al. That is, they should yield a main effect of
Prime Type (active vs. passive), with participants producing more passive sentences
following passive primes, but no effect of Target Verb Type (AP/TE/ET) and no
interaction; as Messenger et al. reported in their study, if the priming effect is
underpinned by abstract knowledge of syntax, it should apply across the board. The
size of the priming effect should also be equivalent for adults and children (i.e. no
interaction of Prime Type by Age), as knowledge of the passive is considered to be
fully adult-like even for young children. If the early semantic restriction account is
correct, our results should show an interaction of Age by Target Verb Type such that
children produce more passives with AP and TE than ET target verbs, while adults
should produce an equal number of passives with all three verb types. This account
also predicts a three-way interaction of Age by Prime Type by Target Verb type,
such that passive priming occurs for all three target verb types for adults, but only
AP and TE target verbs for children.
If the semantic construction prototype account is correct, our results should
show main effects of both (a) Prime Type (i.e. more passives after passive primes)
and (b) Target Verb Type (i.e. more passives for AP and TE than ET target verbs),
but no interaction. This is because the account assumes that knowledge of the passive
is (a) abstract from an early age (hence the main effect of Prime Type and absence of
an interaction with Target Verb Type) but (b) takes the form of a semantic prototype
construction, even for adults (hence the main effect of Target Verb Type). This
account neither predicts nor rules out interactions of Prime Type by Age and Verb
Type by Age, as it is agnostic as to whether or not children’s representations are fully
adult-like by age 4-6. The crucial prediction is that, any such interactions
notwithstanding, the main effects of both (a) Prime Type and (b) Verb Type should
hold for both children and adults when analysed separately. This is because, for both
children and adults, knowledge of the passive (a) is abstract and (b) takes the form of
a semantic prototype construction.
As an addition test of the semantic construction prototype account, and
following Ambridge, Bidgood, Pine, Rowland and Freudenthal (2015), we use
semantic ratings to create a continuous semantic variable, here termed Affectedness,
which is hypothesised to be at the core of the passive semantic construction
prototype. This variable was first used in Ambridge, Bidgood, Pine, Rowland and
134
Freudenthal (2015; labelled as A affects B in that paper). This allows us to move
away from Pinker-style (1989) semantic classes and to test a more fine-grained
approach to semantics. As with the analysis outlined above, if the semantic prototype
account is correct, we predict that this additional analysis will show main effects of
(a) Prime Type (more passives after passive primes) and (b) Semantics (more
passives with verbs rated as having semantics more related to the semantic property
of affectedness), but no interaction. Again, we make no prediction about interactions
of Prime Type by Age and Semantics by Age, although each of the predicted main
effects should hold for both adults and children when analysed separately.
5.2 Method
5.2.1 Participants
The participants were 60 children aged 4-6 years old (4;2-6;5, M=5;6) and 60 adults
aged 19-24. The children were recruited from primary schools and nurseries in the
North West of England and the adults were all undergraduate students at the
University of Liverpool. All participants were monolingual speakers of English and
had no known language impairments.
5.2.2 Test items
Prime verbs consisted of 24 basic agent-patient (AP) verbs (e.g. hug). Target verbs
consisted of 12 AP verbs (e.g. hug), 12 theme-experience (TE) verbs (e.g. annoy)
and 12 experiencer-theme (ET) verbs (e.g. see); all verbs used by Messenger et al.
(2012) were included in our set. All prime and target verbs are given in Table 5.1.
Eight different playlists were created, each of 36 trials. The eight original playlist
orders were reversed to create a further 8 playlists in order to avoid order effects, as a
pilot study suggested that rate of production of passive sentences increased over
time. Prime sentences used 18 different AP verbs, each of which appeared in both an
active and a passive sentence, on separate trials. The remaining 6 AP verbs were
used as targets, along with 6 each of the TE and ET verbs. Participants in the action
were one male and one female character from popular TV animation shows, chosen
to be familiar to young children. Playlists were pseudo-randomised such that no more
than 2 sentences of the same type (active/passive) or two verbs of the same type
135
(AP/TE/ET) appeared in a row. Active and passive sentences containing the same
verb were never used in consecutive trials. The prime sentence always contained
different participants to the target sentence, in order to minimise lexical overlap.
Prime verbs Target verbs
all AP AP ET TE
avoid hold bite* forget amaze
bite* hug carry* hate annoy*
call kick chase hear* bother
carry* kiss dress ignore* frighten*
chase lead hit* know impress
cut pat* hug like* please
dress pull* kick love* scare*
drop push pat* remember* shock*
eat shake pull* see* surprise*
follow squash* push smell tease
help teach squash* understand upset*
hit* wash wash watch worry
Table 5.1. Verbs used as primes and targets. AP = agent-patient verb, ET =
experiencer theme verb, TE = theme-experiencer verb. * denotes verbs also used in
Messenger et al. (2012).
For each prime and target verb, animations were created, using Anime Studio
Pro 5.5, to depict the action. The same animation was used for both the active and
passive versions of prime and target sentences.
5.2.3 Procedure
Syntactic priming was used to encourage experimental participants to produce
passive sentences. Experimenter 1 took turns with the participant to describe
animations presented on a computer screen, using Processing 2
(www.processing.org). A second experimenter, who was unable to see the screen,
gave ‘clue words’ (the prime/target verbs) to Experimenter 1 and the participant.
Experimenter 2 noted down participants’ responses, although sessions were also
audio-recorded, using Audacity, as a backup. Following Rowland et al. (2012), we
used a ‘bingo game’ to motivate the participants to produce responses. For each
sentence produced by Experimenter 1 or the participant, Experimenter 2 looked for a
bingo card that matched the sentence. In fact, Experimenter 2 had all of the bingo
136
cards, but the game was fixed so that the participant always won the game. As the
playlists were long, they were divided into 4 ‘games’ of 9 trials, with participants
requiring 6 bingo cards to win. Before starting, the game was introduced to
participants with three practice trials, using verbs that were not included in the
experiment proper in active locative sentences (e.g. Homer poured water into the
cup).
5.3 Results
Participants’ responses were coded for sentence type, irrespective of prime condition:
correct active, correct (full) passive, incorrect active (with participants reversed),
incorrect (full) passive (with participants reversed), other use of the verb, and
excluded (target verb not used/no response). A response was coded as a correct
active if it was an accurate description of the event, and contained both a subject and
direct object bearing the appropriate role (agent/patient/theme/experiencer) and the
target verb. A response was coded as a correct passive if it was an accurate
description of the event, and contained both a subject and object bearing the
appropriate role (agent/patient/theme/experiencer), an auxiliary verb (get or be), the
target verb and the preposition by. These criteria are similar to those used by
Messenger et al. (2012), with the exceptions that (i) participants in the current study
were required to use the target verb, and (ii) the range of semantic roles was more
varied, as we used AP, ET and TE verbs as targets, whereas Messenger et al. used
only AP verbs as targets. Table 5.2 shows the frequency of each response type for
both adults and children.
137
Correct
active
Correct
passive
Incorrect
(reversed)
active
Incorrect
(reversed)
passive
Other
use of
verb Excluded
Adults 1544 424 41 14 80 57
Children 1355 133 53 6 136 477
Table 5.2. Number of responses of each type by age group.
Again following Messenger et al., the dependent variable in our analysis was
binary (correct active or correct passive response, with all other responses excluded).
Results were analysed in RStudio (version 0.98.1103; R version 3.2.0, R Core Team,
2015). As the dependent variable was binary, results were analysed using the glmer
function of the lme4 package (version 1.1-7, Bates, Maechler, Bolker & Walker,
2014). Predictor variables were Age Group (adult/child), Prime Type (active/passive)
and Target Verb Type (AP/TE/ET). Factors were centred prior to analysis and
Helmert contrast coding was used for the 3-way factor of Target Verb Type. Random
intercepts for Verb and Participant were included in the model, although no random
slopes were included as their inclusion prevented the model from converging.
Interactions for Age by Prime Type and Age by Verb Type were included in the
model, although no other interactions were included as this also prevented the model
from converging. Because we had to exclude the three-way interaction of Age by
Prime Type by Target Verb Type, we ran separate models for adults and children in
order to test the prediction that passive priming occurs for all three target verb types
for adults, but only AP and TE target verbs for children. A plot of mean passive
responses, by Age Group, Prime Type and Target Verb Type, is shown in Figure 5.1.
138
Figure 5.1. Proportion of correct passives produced, by age group (adult/child),
prime type (active/passive) and target verb type (theme-experiencer, e.g. scare /
agent-patient, e.g. hug / experiencer-theme, e.g. see).
Table 5.3 shows the model summary, with model comparison results shown
in Table 5.4. Model comparisons revealed a main effect of Age Group
(χ2[df=1]=23.20, p<0.001), such that adults (M=0.22, SE=0.01) produced more passive
sentences than children (M=0.09, SE=0.01). This effect was not specifically
predicted by any of the three accounts under investigation, but is unsurprising given
that passives are relatively difficult in general (i.e. they are longer than actives, and
reverse active word order). Consistent with both the early abstraction and semantic
construction prototype accounts, model comparisons also revealed a main effect of
Prime Type (χ2[df=1]=101.05, p<0.001), such that participants produced more passive
sentences following passive (M=0.22, SE=0.01) than active primes (M=0.11,
SE=0.01). The interaction between Age Group and Prime Type was non-significant,
as predicted by the early abstraction account, but the p value was .061, so may have
become significant with more power (χ2[df=1]=3.51, p=0.061, n.s.). Therefore, on a
conservative reading, the prediction of the early abstraction account here only
receives tentative support.
139
Fixed effects
Estimate (SE)
Intercept -2.92 (0.19)
Age Group 1.59 (0.33)
Prime Type 1.23 (0.13)
Target Verb Type (affected subject; AP+TE
vs. ET)
1.67 (0.23)
Target Verb Type (action; AP vs. TE) -0.78 (0.22)
Age Group x Prime Type -0.53 (0.27)
Age Group x Target Verb Type (affected
subject; AP+TE vs. ET)
-0.17 (0.35)
Age Group x Target Verb Type (action; AP
vs. TE)
0.69 (0.27)
Random effects
Variance (SD)
Participant (intercept) 1.72 (1.31)
Target Verb (intercept) 0.19 (0.43)
Model summary AIC 2433.3
BIC 2494.8
Log Likelihood -1206.6
Deviance 2413.3
Table 5.3. Model summary for all participants.
140
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 3 2594.2 2612.6 -1294.1 2588.2
Age Group 4 2573.0 2597.6 -1282.5 2565.0 23.20 1 1.46e-6
***
Prime Type 5 2473.9 2504.7 -1232.0 2463.9 101.05 1 <2.2e-16
***
Target Verb
Type (all)
7 2437.4 2480.4 -1211.7 2423.4 40.51 2 1.60e-9
***
Target Verb
Type
(affected
subject;
AP+TE vs.
ET)
6 2442.8 2479.6 -1215.4 2430.8 33.182 1 8.39e-9
***
Target Verb
Type
(action; AP
vs. TE)
7 2437.4 2480.4 -1211.7 2423.4 7.33 1 0.0068
**
Age Group
x Prime
Type
8 2435.9 2485.1 -1210.0 2419.9 3.51 1 0.061,
n.s.
Age Group
x Target
Verb Type
(all)
10 2433.3 2494.8 -1206.7 2413.3 6.62 2 0.036 *
Age Group
x Target
Verb Type
(affected
subject;
AP+TE vs.
ET)
9 2437.7 2493.0 01209.8 2149.7 0.25 1 0.62,
n.s.
Age Group
x Target
Verb Type
(action; AP
vs. TE)
10 2433.3 2494.8 -1206.7 2413.3 6.37 1 0.012 *
Table 5.4. Model comparisons for all participants.
Importantly, and in support of the semantic construction prototype account,
model comparisons revealed a main effect of Target Verb Type (χ2[df=2]=40.51,
p<0.001). This is also compatible with the early semantic restriction account,
provided that the main effect is driven by differences in the children’s responses, and
not those of the adults (see separate analyses, below). Contrasts revealed that
participants produced significantly more passive sentences with AP and TE verbs
141
(verbs involving passive subject affectedness) than with ET verbs (M=0.07,
SE=0.01) (χ2[df=1]=33.18, p<0.001) and significantly more passive sentences with TE
(M=0.26, SE=0.01) than AP verbs (M=0.17, SE=0.01) (χ2[df=1]=7.33, p=0.007).
The early semantic restriction account (see section 5.2.1) predicts an
interaction of Age Group by Target Verb Type such that children will produce more
passives with AP and TE than ET target verbs, while adults will produce an equal
number of passives with all three types. Although a significant interaction was
observed (χ2[df=2]=6.62, p=0.037), inspection of Figure 5.1 reveals that it does not
conform to the pattern predicted by this account. Rather, the pattern of by-verb
differences was similar for the two age groups. In order to unpack this interaction,
we conducted further analyses for each age group separately. These separate analyses
also allow for the investigation of the prediction of the early semantic restriction
account that passive priming will occur for all three target verb types for adults, but
only AP and TE target verbs for children.
5.3.1 Adults
Results were analysed using linear mixed models, as above. A by-participant random
slope for Prime Type was included in the model, although the model would not
converge with additional by-participant random slopes (by-verb random slopes were
not meaningful, given the design). A Prime Type by Target Verb Type interaction
was also included. Table 5.5 shows the model summary, with model comparison
results shown in Table 5.6. Model comparisons revealed a main effect of Prime Type
(χ2[df=1]=21.26, p<0.001), such that adults produced more passive sentences following
passive (M=0.28, SE=0.01) than active primes (M=0.15, SE=0.01). Model
comparisons also revealed a main effect of Target Verb Type (χ2[df=2]=38.51,
p<0.001). Contrasts revealed that adults produced significantly more passive
sentences with AP and TE verbs (verbs involving passive-subject affectedness) than
with ET verbs (M=0.09, SE=0.01) (χ2[df=1]=34.83, p<0.001) as well as, unexpectedly,
significantly more passive sentences with TE (M=0.31, SE=0.02) than AP verbs
(M=0.24, SE=0,02) (χ2[df=1]=3.92, p=0.048)1. No Prime Type by Target Verb Type
1 This result may seem surprising, given that previous research has generally assumed prototypical
passives to be actional AP verbs. However, it is worth noting that events denoted by TE verbs are, by
definition, highly affecting for the passive subject (see Ambridge et al., 2015). In addition, Maratsos
142
interaction was observed (χ2[df=2]=0.42, p=0.81, n.s.), suggesting that Prime Type
affected adults’ responses in the same way for all target verb types. Thus, in
summary, the results for adults pattern as predicted by the semantic prototype
account, but not the early semantic restriction or early abstraction accounts (neither
of which predicts Target Verb effects in adults), with main effects of both (a) Prime
Type (i.e. more passives after passive primes) and (b) Target Verb Type (i.e. more
passives for AP and TE than ET target verbs), but no interaction.
Fixed effects
Estimate (SE)
Intercept -2.18 (0.20)
Prime Type 1.10 (0.24)
Target Verb Type (affected subject; AP+TE
vs. ET)
1.64 (0.22)
Target Verb Type (action; AP vs. TE) -0.43 (0.21)
Prime Type x Target Verb Type (affected
subject; AP+TE vs. ET)
-0.17 (0.33)
Prime Type x Target Verb Type (action; AP
vs. TE)
-0.13 (0.28)
Random effects
Variance (SD)
Participant (intercept) 1.46 (1.21)
Participant – Prime Type 1.09 (1.04)
Target Verb (intercept) 0.15 (0.39)
Model
summary
AIC 1729.6
BIC 1785.5
Log Likelihood -854.8
Deviance 1709.6
Table 5.5. Model summary for adults.
et al. (1985) suggested that input frequency of the verb types in question may have an influence on
passive production, with change-of-state verbs, such as TE verbs, being of higher frequency in the
passive in child-directed speech than verbs describing temporary physical contact, such as kick, and
many of the other AP verbs used in the current study.
143
Predictor df AIC BIC logLik deviance χ2 χ
df
p
value
Intercept 5 1780.1 1808.0 -885.03 1770.1
Prime Type 6 1760.1 1794.3 -874.40 1748.8 21.26 1 1.02e-6
***
Target Verb
Type (all)
8 1726.1 1770.7 -855.03 1710.1 38.75 2 3.85e-9
***
Target Verb
Type
(affected
subject;
AP+TE vs.
ET)
7 1728.0 1767.1 -856.99 1714.0 34.83 1 3.60e-9
***
Target Verb
Type (action;
AP vs. TE)
8 1726.1 1770.7 -855.03 1710.1 3.92 1 0.048
*
Prime Type x
Target Verb
Type (all)
10 1729.6 1785.5 -854.82 1709.6 0.42 2 0.81,
n.s.
Prime Type x
Target Verb
Type
(affected
subject;
AP+TE vs.
ET)
9 1727.8 1778.1 -854.91 1709.8 0.24 1 0.63,
n.s.
Prime Type x
Target Verb
Type (action;
AP vs. TE)
10 1729.6 1785.5 -854.82 1709.6 0.19 1 0.67,
n.s.
Table 5.6. Model comparisons for adults.
5.3.2 Children
Results were analysed using linear mixed models as for adults, above, except that no
random slopes were included, as this prevented the model from converging. Table
5.7 shows the model summary, with model comparison results shown in Table 5.8.
Model comparisons revealed a main effect of Prime Type (χ2[df=1]=48.54, p<0.001),
such that children produced more passive sentences following passive (M=0.14,
SE=0.01) than active primes (M=0.04, SE=0.01). Model comparisons also revealed a
main effect of Target Verb Type (χ2[df=2]=22.19, p<0.001). Contrasts revealed that
children produced significantly more passive sentences with AP and TE verbs (verbs
involving passive-subject affectedness) than with ET verbs (M=0.03, SE=0.01)
144
(χ2[df=1]=14.19, p<0.001) and, again unexpectedly, significantly more passive
sentences with TE (M=0.18, SE=0.02) than AP verbs (M=0.08, SE=0.02)
(χ2[df=1]=8.00, p=0.005). No Prime Type by Target Verb Type interaction was
observed (χ2[df=2]=0.15, p=0.93, n.s.), suggesting that Prime Type affected children’s
responses in the same way for all target verb types. Thus the results for children also
pattern as predicted by the semantic prototype account, but not the early semantic
restriction or early abstraction accounts, with main effects of both (a) Prime Type
(i.e. more passives following passive primes) and (b) Target Verb Type (i.e. more
passives for AP and TE than ET target verbs), but no interaction.
Fixed effects
Estimate (SE)
Intercept -3.95 (0.37)
Prime Type 1.55 (0.37)
Target Verb Type (affected subject; AP+TE
vs. ET)
1.83 (0.46)
Target Verb Type (action; AP vs. TE) -1.16 (0.41)
Prime Type x Target Verb Type (affected
subject; AP+TE vs. ET)
0.04 (0.73)
Prime Type x Target Verb Type (action; AP
vs. TE)
-0.19 (0.52)
Random effects
Variance (SD)
Participant (intercept) 2.43 (1.56)
Target Verb (intercept) 0.55 (0.74)
Model summary AIC 695.2
BIC 737.6
Log Likelihood -339.6
Deviance 679.2
Table 5.7. Model summary for children.
145
Predictor df AIC BIC logLik deviance χ2 χ
df
p value
Intercept 3 756.05 771.96 -375.02 750.05
Prime
Type
4 709.51 730.73 -350.75 701.51 48.54 1 3.24e-12
***
Target
Verb Type
(all)
6 691.32 723.15 -339.66 679.32 22.19 2 1.52e-5
***
Target
Verb Type
(affected
subject;
AP+TE
vs. ET)
5 697.32 723.85 -343.66 687.32 14.19 1 1.7e-5
***
Target
Verb Type
(action;
AP vs.
TE)
6 691.32 723.15 -339.66 679.32 8.00 1 0.0047
**
Prime
Type x
Target
Verb Type
(all)
8 695.17 737.61 -339.59 679.17 0.15 2 0.93,
n.s.
Prime
Type x
Target
Verb Type
(affected
subject;
AP+TE
vs. ET)
7 693.31 730.45 -339.66 679.31 0.005
7
1 0.94,
n.s.
Prime
Type x
Target
Verb Type
(action;
AP vs.
TE)
8 695.17 737.61 -339.59 679.17 0.14 1 0.71,
n.s.
Table 5.8. Model comparisons for children.
5.3.3 Fine-grained semantic analysis
In order to test a fine-grained approach to semantics, we re-ran the analyses above
but replaced the discrete semantic classes with a continuous semantic variable,
previously used in Ambridge, Bidgood, Pine, Rowland and Freudenthal (2015). The
146
variable was derived by collecting semantic ratings for each verb from ten adults
(who did not take part in the priming study) for ten semantic features thought to be
associated with the passive construction (e.g. the agent is doing something to the
patient, the agent is responsible). Principal components analysis was then used to
derive a single, composite variable. This variable is named Affectedness, as the
biggest weightings are related to the extent to which the subject in an active sentence
affects the object (irrespective of their semantic roles). As with the other predictor
variables, this variable was centred before being entered into mixed effects models
using the glmer function in R. The outcome variable was the proportion of correct
passive responses out of the total of correct active and correct passives.
Predictor variables were Age Group (adult/child), Prime Type
(active/passive) and Semantics (Affectedness). (Predictor variables were centred prior
to analysis.) Random intercepts for Verb and Participant were included in the model,
although no random slopes were included, as this prevented the model from
converging. Interactions for Age by Prime Type and Age by Semantics were
included in the model, although the three-way interaction was removed as this also
prevented the model from converging. Table 5.9 shows the model summary, with
model comparison results shown in Table 5.10.
Fixed effects Estimate (SE) Intercept -2.60 (0.20)
Age Group 1.44 (0.30)
Prime Type 1.25 (0.14)
Semantics (Affectedness) 0.47 (0.11)
Age Group x Prime Type -0.48 (0.26)
Age Group x Semantics 0.19 (0.12)
Prime Type x Semantics -0.10 (0.11)
Random effects Variance (SD) Participant (intercept) 1.70 (1.30)
Target Verb (intercept) 0.50 (0.70)
Model summary AIC 2457.7
BIC 2513.0
Log Likelihood -1219.8
Deviance 2439.7
Table 5.9. Model summary for all participants.
147
Predictor df AIC BIC logLik deviance χ2 χ
df
p
value
Intercept 3 2594.2 2612.6 -1294.1 2588.2
Age
Group
4 2573.0 2597.6 -1282.5 2565.0 23.2 1 1.5e-6
***
Prime
Type
5 2473.9 2504.7 -1232.0 2463.9 101.05 1 >2.2e-
16 ***
Semantics
(Affectedn
ess)
6 2459.1 2496.0 -1223.5 2447.1 16.838 1 4.1e-5
***
Age
Group x
Prime
Type
7 2457.5 2500.6 -1221.8 2443.5 3.57 1 0.059,
n.s.
Age
Group x
Semantics
8 2456.5 2505.7 -1220.2 2440.5 3.026 1 0.082,
n.s.
Prime
Type x
Semantics
9 2457.7 2513.0 -1219.8 2439.7 0.826 1 0.36,
n.s.
Table 5.10. Model comparisons for all participants.
In support of the semantic construction prototype account, and in line with
the results of our initial analysis, model comparisons revealed a main effect of
Semantics (χ2[df=1]=16.84, p<0.001). Figure 5.2 shows that, the higher a verb is rated
on the Affectedness scale, the more likely participants are to produce a correct
passive sentence with that verb. The pattern is the same for both Age Groups, and for
both Prime Types. As with the model presented in Table 5.3, main effects of Age
Group (χ2[df=1]=23.20, p<0.001) and Prime Type (χ2
[df=1]=101.05, p<0.001) were also
observed. No significant interactions were found, although two marginal interactions
were observed (Age Group by Prime Type: χ2[df=1]=3.57, p=0.059; Age Group by
semantics: χ2[df=1]=3.03, p=0.082). Figure 5.2 indications that these marginal
interactions may be driven by children’s responses when primed with active
sentences, which appear to be close to a floor effect.
148
Figure 5.2. Proportion of correct passives plotted against semantic rating
(Affectedness), split by Prime Type and Age Group.
As no Age Group by Semantics interaction was observed, the findings are
incompatible with the predictions of the early semantic restriction account, as this
account predicts that by-verb differences should only be observed for children.
Nevertheless, as in the initial analysis, we ran analyses with each age group
separately. Table 5.11 shows the model for summary for adults, with model
comparisons for this age group shown in Table 5.12. Table 5.13 shows the model for
summary for children, with model comparisons for this age group shown in Table
5.14. In both of these models, a random slope for Prime Type was included, although
all other random slopes were removed to enable the model to converge. Main effects
for Prime Type and Semantics (Affectedness) were observed for both Age Groups,
but there was no interaction for either age group. These results again support the
early abstraction account, with young children demonstrating abstract knowledge of
the passive, and the semantic prototype account, with adults and children both
increasingly likely to produce a passive sentence with verbs more compatible with
the semantic feature of Affectedness. These results to not support the early semantic
restriction account, as adults and children appear to be using semantic information in
the same way.
149
Fixed effects Estimate (SE) Intercept -1.90 (0.21)
Prime Type 1.11 (0.22)
Semantics (Affectedness) 0.57 (0.11)
Prime Type x Semantics -0.17 (0.12)
Random effects Variance (SD) Participant (intercept) 1.47 (1.21)
Participant (Prime Type) 1.11 (0.06)
Target Verb (intercept) 0.36 (0.60)
Model summary AIC 1740.7
BIC 1785.4
Log Likelihood -862.3
Deviance 1724.7
Table 5.11. Model summary for adults.
Predictor df AIC BIC logLik deviance χ2 χ
df
p
value
Intercept 5 1780.1 1808.0 -885.03 1770.1
Prime
Type
6 1760.8 1794.3 -874.40 1748.8 21.255 1 4.0e-6
***
Semantics
(Affectedn
ess)
7 1740.4 1779.5 -863.20 1726.4 22.416 1 2.2e-6
***
Prime
Type x
Semantics
8 1740.7 1785.4 -862.35 1724.7 1.6922 1 0.19,
n.s.
Table 5.12. Model comparisons for adults.
Fixed effects Estimate (SE) Intercept -3.87 (0.45)
Prime Type 2.18 (0.62)
Semantics (Affectedness) 0.38 (0.21)
Prime Type x Semantics 0.09 (0.24)
Random effects Variance (SD) Participant (intercept) 3.03 (1.74)
Participant (Prime Type) 0.38 (0.61)
Target Verb (intercept) 1.32 (1.15)
Model summary AIC 711.73
BIC 754.17
Log Likelihood -347.86
Deviance 695.73
Table 5.13. Model summary for children.
150
Predictor df AIC BIC logLik deviance χ2 χ
df
p
value
Intercept 5 732.25 758.78 -361.13 722.25
Prime
Type
6 711.78 743.62 -349.89 699.78 22.466 1 2.1e-6
***
Semantics
(Affectedn
ess)
7 709.87 747.00 -347.93 695.73 3.9173 1 0.048
*
Prime
Type x
Semantics
8 711.73 754.17 -347.86 695.73 0.1411 1 0.71,
n.s.
Table 5.14. Model comparisons for children.
5.3.4 Summary
Despite a significant Age Group by Target Verb Type interaction in the initial
analysis, the pattern of results for adults and children is remarkably similar, as
illustrated in Figure 5.1. This suggests that the pattern of results per se does not differ
by age group but, rather, that the interaction reflects a difference in magnitude
between the proportion of passives produced with the three different target verb
types in the two age groups, particularly between TE and AP verbs. In addition, no
Age by Semantics interaction was observed in the fine-grained semantic analysis.
Thus, verb semantics appears to affect the production of passives for children and
adults in the same way.
5.4 Discussion
Using production priming, the current study has provided additional support for
Messenger et al.’s (2012) claim that both adults and young children have abstract
syntactic knowledge of the passive, whilst our reversal of Messenger et al.’s design
has highlighted the fact that, nevertheless, by-verb differences do exist. Our findings
therefore support Ambridge, Bidgood, Pine, Rowland and Freudenthal’s (2015)
claim that adults’ abstract knowledge of the passive takes the form of a semantic
construction prototype, and add weight to this account by demonstrating that this
holds for production and is also already the case for young children. The findings are
not consistent with claims that children’s acquisition of passive syntax is delayed
(Borer & Wexler, 1987; Fox & Grodzinsky, 1998), nor that children’s knowledge is
151
restricted to a core of actional verbs (e.g. Maratsos et al., 1985), at least not at the age
of 4 years.
Thus, as we suggested in the introduction, the early semantic restriction
account is right in highlighting by-verb semantic differences in the passive (although,
in fact, these persist into adulthood) while the early abstraction account is right in
highlighting the fact that, nevertheless, both adults and children have an abstract
verb-general representation of the construction (though one that does not preclude
by-verb semantic differences). Rather, we need an account that captures the insights
of both of these accounts, explaining both early abstract knowledge and late by-verb
semantic differences. The semantic construction prototype account is one such
account, but this raises the question of exactly what it means to have an abstract
construction that, nevertheless, constitutes a semantic prototype.
In fact, this scenario is one that is familiar in cognitive psychology (e.g.
Rosch, 1975). For example, humans have an abstract concept of “bird” that they have
formed (presumably) by generalising across concrete instances of this category that
they have witnessed. Although this category is abstract (the ‘bird’ prototype may not
be a real bird but an amalgamation of typical features in the birds people have
encountered, c.f. exemplar theory, e.g. Nosofsky, 1986), it nevertheless has a
prototype structure. People have no difficulty in recognising novel species of bird.
However, prototypical instances (e.g. robin) enjoy an advantage over less
prototypical instances (e.g. ostrich) in (a) judgment, (b) reaction time and (c)
production priming tasks. What makes robin a more prototypical member of the
category than ostrich is the fact that it shares greater overlap with other category
members along the dimensions that are relevant to category membership (e.g. size,
ability to fly). But, importantly, ostrich still shares enough similarities with other
members (e.g. it has wings and a beak and lays eggs) to be included in the category.
In the same way, English speakers appear to have an abstract concept of the
passive construction that they have formed by generalising across concrete instances
to which they have been exposed. Although this construction is abstract (for
example, people have no difficulty using novel verbs in this construction; Pinker et
al., 1987), it nevertheless has a prototype structure (which we are capturing with the
label “affectedness”). (See Ibbotson & Tomasello, 2009, for a discussion of how
prototypes in language may lead to the formation of abstract schemas and
constructions.) Prototypical instances (e.g. passives with kick [AP] and frighten [TE])
152
enjoy an advantage over less prototypical instances (e.g. passives with see [ET]) in
(a) judgment (Ambridge, Bidgood, Pine, Rowland & Freudenthal, 2015), (b) reaction
time (Ambridge, Bidgood, Pine, Rowland & Freudenthal, 2015) and (c) production
priming tasks (the present study). What makes passives with AP and TE verbs more
prototypical members of the category than passives with ET verbs is the fact that the
former share greater overlap with other category members along the dimensions that
are relevant to category membership (e.g. semantic affectedness). But, importantly,
passives with ET verbs still share enough similarities with other members (i.e. they
include a by-phrase and passive morphology, reverse canonical linking and
discourse-focus patterns) to be included in the category.
In order to compare the semantic verb class account with the fine-grained
semantic analysis we performed, we plotted the proportion of passives produced by
adults and children against the semantic continuum of Affectedness (Figure 5.3). The
colours indicate the three original verb classes (AP, ET and TE). The Figure shows
that AP and TE verbs overlap, both in terms of their semantic ratings and the
proportion of passive sentences the children and adults in the study produced with
those verbs. Perhaps the traditional distinction between these classes is therefore not
a particularly meaningful one. In contrast, the ET verbs seem to form a cluster in the
low-affectedness/low-passive-production quartile of the Figure. Nevertheless,
particularly for adults, the figure shows that verbs of this type vary noticeably in
terms of the proportion of passive sentences produced based on the verb’s
Affectedness rating. Overall, then, meaningful by-verb differences in semantics
seems to be best captured using a continuum, rather than discrete classes. This
supports the idea of a prototypical structure, with prototypical passive sentences
containing verbs that are highly rated in terms of Affectedness, such as frighten and
hit.
153
Figure 5.3. Proportion of correct passives out of correct actives and passives
combined produced by adults and children, plotted against Affectedness rating.
Colours indicate the traditional semantic classes.
Of course, the idea that linguistic categories have a semantic prototype
structure is not a new one (e.g. Langacker, 1987; Goldberg, 2006; Dąbrowska,
Rowland & Theakston, 2009). Although, until recently, there had been little
supporting evidence for this view in the domain of verb argument structure
constructions, the present study joins a growing body of research finding exactly
such effects. For example, in a grammaticality judgment study, Ambridge, Pine and
Rowland (2012) found evidence supporting Pinker’s (1989) claim that the
prototypical semantics of the ground locative construction (e.g. Homer filled the cup
with water) involve end-state, whereas the prototypical semantics of the figure
locative construction (e.g. Homer poured water into the cup) involve manner of
motion. Adult participants’ grammaticality judgments were significantly predicted by
the extent to which verb semantics were consistent with the end-state manner-of-
motion semantics for ground- and figure-locative sentences (as rated by independent
participants), respectively. Children (aged 5 and 9 years) also showed this effect but
for figure-locative sentences only, suggesting development of locative construction
semantic knowledge continues beyond 9 years of age (see also Bidgood et al., 2014).
In a similar grammaticality judgment study, Ambridge, Pine, Rowland,
Freudenthal and Chang (2014) demonstrated the psychological reality of semantic
154
constraints on the dative constructions, again based on Pinker (1989). The
prototypical semantics of the double-object (DO) dative construction (e.g. Bart threw
Lisa the ball) relate to causing to have, whereas the prototypical semantics of the
prepositional-object (PO) dative construction (e.g. Bart threw the ball to Lisa) relate
to causing to go (in a particular manner). Both adults and children rated verbs with
semantics consistent with the prototypical semantics of the construction in which
they appeared as more acceptable (see Ambridge, 2013; Blything et al., 2014, for
similar findings for the un-VERB reversative construction).
More generally, the findings of the present study suggest that the familiar
dichotomy between early-abstraction accounts and those that posit a stage
characterised by lexical and semantic restrictions is too simplistic. Instead, these
findings suggest the need for an account of acquisition that combines the insights of
both approaches. Yes, children rapidly abstract across concrete exemplars to acquire
abstract representations relatively early in development, but these abstract
representations nevertheless retain the character of the original exemplars that gave
rise to them. The semantic prototype theory constitutes one way (but not necessarily
the only way) of explaining these findings.
Setting aside these broader considerations, in conclusion, the current paper
suggests that, by the age of 4 years, children have an abstract underlying syntactic
representation of the passive. This knowledge, however, has semantic structure that
persists into adulthood.
5.5 Acknowledgements
This work was supported by the ESRC International Centre for Language and
Communicative Development (LuCiD). The support of the Economic and Social
Research Council [ES/L008955/1] is gratefully acknowledged. This work was also
supported by Grant RPG-158 from the Leverhulme Trust. We would also like to
thank the schools, teachers, parents and children who made this research possible,
and the undergraduate students who assisted with data collection.
155
Chapter 6: Discussion
6.1 Introduction
This thesis set out to investigate children’s argument structure overgeneralisation
errors (e.g. *Don’t giggle me!) and the mechanisms that have been proposed to
explain how children ‘retreat’ from such errors, or, indeed, avoid making them
altogether. Due to the prevalence of lexical effects reported in the literature, the
thesis was based on a constructivist framework. Through a series of experiments,
three approaches to explaining children’s retreat from argument structure
overgeneralisation errors were examined: semantics (e.g. Pinker, 1989),
entrenchment (Braine & Brookes, 1995) and preemption (e.g. Goldberg, 1995).
Chapter 3 presented a critical test of the semantics and frequency accounts: a novel-
verb grammaticality judgment study of the locative alternation. Chapter 4 examined
the most frequent alternation – the causative – using both grammaticality judgment
and error elicitation tasks. Chapter 5 extend the investigation of the role of semantics
beyond argument structure overgeneralisation errors to the passive construction, with
the aim of testing whether semantics might play a wider role in children’s acquisition
of argument structure.
This Discussion chapter will first summarise the findings of the three
experimental chapters presented in this thesis. It will discuss how these results
contribute to the growing body of evidence that, although individual mechanisms
each enjoy some support, children’s retreat from overgeneralisation errors is best
explained by a theory that integrates elements of the entrenchment, preemption and
semantic accounts. The chapter concludes by suggesting that future research should
focus on testing integrated accounts of argument structure acquisition, such as
Ambridge et al.’s FIT account (Ambrige & Lieven, 2011; Ambridge et al., 2011).
6.2 What do the studies in this thesis tell us about the retreat from
overgeneralisation?
Overgeneralisation errors, such as *Don’t giggle me!, are thought to stem from
children noticing that some verbs can appear in two argument structures with similar
meaning, e.g. Christine sprayed water onto the flowers and Christine sprayed the
156
flowers with water. As various different verbs appear in the same distribution in the
input (e.g. splash, load), children are able to usefully generalise this pattern to new
verbs (e.g. sprinkle, pack). If children apply this pattern too widely, to verbs that can
only appear in one or other of the constructions (e.g. pour, fill), overgeneralisation
errors will result, such as *Howard poured the cup with tea/filled tea into the cup.
This thesis began by investigating how three proposed mechanisms were able to
explain the retreat from such overgeneralisation errors: semantics, based initially on
the semantic verb class hypothesis (Pinker, 1989); entrenchment (Braine & Brooks,
1995); and preemption (Goldberg, 1995).
According to Pinker’s semantic verb class hypothesis, verbs with similar
semantics are grouped into classes (e.g. ‘locative verbs’) and more specific
subclasses (e.g. one in which “a mass is enabled to move via the force of gravity”).
Each of the subclasses may be alternating or non-alternating, but children who are
yet to construct these subclasses may erroneously treat all verbs as alternating. Not
until children have correctly set up these subclasses will they be able to avoid, or
retreat from, overgeneralisation errors.
According to the entrenchment hypothesis, children may make
overgeneralisation errors when they have not received enough information in the
input to rule out the possibility that certain verbs may not alternate between
competing constructions. By frequently hearing verbs in grammatical sentences,
children are eventually able to infer from their absence in the input that certain verb-
construction pairings must not be acceptable.
Finally, like the entrenchment hypothesis, the preemption hypothesis
proposes that lack of evidence in the input can lead to overgeneralisation errors.
However, this hypothesis proposes that only hearing verbs used in constructions with
very similar semantics to the ungrammatical construction will lead children to infer
that the latter must be unacceptable, and thus lead to the retreat from
overgeneralisation.
Chapter 3 used the locative alternation (Christine sprayed water onto the
flowers/sprayed the flowers with water) to test the predictions of the semantic verb
class hypothesis and the ‘frequency’ hypothesis (it was not possible to differentiate
between the predictions of the entrenchment and preemption hypotheses in this
study). Chapter 4 used the causative alternation (John rolled the ball/The ball rolled)
to investigate all three hypotheses separately, although semantics was treated as a
157
continuum (e.g. Ambridge et al., 2014) rather than relying on the discrete classes
proposed by Pinker. Chapter 5 moved beyond overgeneralisation errors to investigate
the predictions of the semantic verb class hypothesis in more detail with a
construction that children find notoriously difficult: the passive (e.g. Bob was helped
by Wendy). Altogether, the experimental evidence presented here provides strong
support for the roles of both verb semantics and verb frequency effects, although the
role played by preemption is less clear than that of entrenchment. Overall, the
conclusion is that none of these mechanisms alone can explain children’s retreat
from overgeneralisation, or their acquisition of argument structure more widely.
Instead, an integrated account, such as the FIT account (Ambridge & Lieven, 2011;
Ambridge et al., 2011), is needed to explain the results presenting here and in
previous papers.
6.2.1 Chapter 3: Locatives
Chapter 3 presented a novel-verb grammaticality judgment study of the locative
alternation (e.g. Christine sprayed water onto the flowers/sprayed the flowers with
water). Adults and children (aged 5-6 and 9-10) were taught novel verbs. These were
presented with animations illustrating their meanings, but were never presented in
locative sentences. Thus, grammaticality judgments of these novel verbs had to be
based on their semantics. The semantics of the verbs were designed to match with
one of three of Pinker’s (1989) fine-grained semantic subclasses of locative verbs –
one ground-only class (containing verbs like fill), one figure-only (containing verbs
like pour) and one alternating class (containing verbs like spray). If participants
judged sentences containing the novel verbs in line with sentences containing real
verbs from the same semantic subclass, then they must be able to use verb semantics
to determine the construction(s) in which those verbs can be used grammatically. To
test the ‘frequency’ hypothesis (it was not possible to distinguish between the
entrenchment and preemption hypotheses in this study), we included high-frequency
and low-frequency real verbs from each of the three semantic subclasses, along with
the novel verbs. If participants use verb frequency information to determine if a verb
is grammatical in a certain construction, they should be less willing to accept
ungrammatical sentences the more frequently that verb has previously been heard.
158
The results of the experiment were analysed using ANOVAs. When looking
at only novel verbs – they key test of the semantic verb class hypothesis in this
chapter – results for all participants followed the predicted pattern for novel non-
alternating verbs, providing strong support for the semantic verb class hypothesis.
Unexpectedly, participants showed an overall preference for the ground-locative uses
of alternating novel spray-type verbs. We speculated that this might be related to the
holism constraint: in order for animations to be compatible with the ground-locative
construction, the whole of the ‘ground’ object needed to be affected, e.g. all of the
roses needed to be sprayed with water for the sentence Lisa sprayed the roses with
water to be felicitous. Therefore, participants could have judged the figure locative,
Lisa sprayed water onto the flowers, to be less acceptable (although ratings were still
broadly grammatical for both sentence types for the alternating verbs).
Strong support was also provided for the frequency account. Results revealed
a greater dispreference for ungrammatical uses of high-frequency non-alternating
verbs (e.g. fill, pour) than for low-frequency (e.g. line, drip) or novel verbs from the
same subclasses. The dispreference for ungrammatical uses of low-frequency non-
alternating verbs was also greater than the dispreference for novel verbs in the same
class, although this difference was much smaller.
There were also differences between the age groups in this study. Unlike the
adults and older children, the 5- to 6-year-olds showed no preference for either
construction type for novel ground-only verbs (e.g. fill). We suggested that the fine-
grained semantics of this verb class might have been too difficult for these young
children to have fully grasped. Only the adults showed the unexpected preference for
ground locative uses of novel alternating verbs. As mentioned above, this could have
been due to the holism constraint: perhaps the complex semantics of this constraint
have not yet been acquired by either group of children in the current study. Adult
participants also showed a greater degree of dispreference for the ungrammatical
sentences for non-alternating verbs than either group of children, and the older
children showed a greater degree of dispreference for the ungrammatical sentences
than did the younger children. This could be interpreted as providing further support
for the frequency hypothesis, since adults have, presumably, had more exposure to
all of the (real) verbs in question in competing constructions than the children have,
and therefore would have been more confident in their judgments of the relative
159
(un)grammaticality of the sentences presented here. All of these by-age differences,
then, could be taken as further support for the hypotheses under investigation.
Overall, the results from Chapter 3 cannot be explained by either the
semantics or frequency accounts alone. An integrated account that can explain both
of these effects seems necessary.
6.2.2 Chapter 4: Causatives
This chapter presented three experiments testing the semantics, entrenchment and
preemption hypotheses, using the causative alternation (e.g. John rolled the ball/The
ball rolled). Experiment 1 was a grammaticality judgment study with adults, using a
large number of verbs. Experiment 2 was another grammaticality judgment study,
but it used a smaller number of verbs in order to enable us to test children as well as
adults (young children would have been unlikely to know all of the verbs used in
Experiment 1 and not all of the verbs would have been suitable for them, e.g.
murder). Experiment 3 employed a novel use of the production-priming method in a
(very successful) attempt to elicit overgeneralisation errors from 5- to 6-year-old
children. Unlike the locatives study in Chapter 3, this chapter used only real verbs. In
addition, rather than testing the semantic verb class hypothesis specifically, we tested
a more general semantic account by collecting semantic ratings from a group of
adults (who did not take part in the main experiments) to create a continuum of
causative semantics. This approach allows for more flexibility to investigate widely-
observed lexical effects, and follows a method similar to that used by Ambridge and
colleagues’ investigations of the locative and dative constructions (Ambridge et al.,
2014, and Ambridge, Pine, Rowland & Chang, 2012, respectively). Corpus counts
were used to test both the entrenchment and preemption hypotheses, with overall
verb frequency used as the predictor variable for the entrenchment hypothesis and
verb frequency in a semantically-related competing construction used to test the
preemption hypothesis (the periphrastic causative for intransitive-only verbs, e.g.
Homer made the fish swim, and the passive for transitive-only verbs, e.g. The ball
was kicked).
In both grammaticality judgment studies, we calculated difference scores for
non-alternating verbs by taking the rating for the ungrammatical sentence away from
the rating for the grammatical sentence. This was important as it allowed us to
160
control for any general preferences for certain animations, or judgments about the
likelihood of certain situations occurring. These scores were calculated separately for
intransitivisation errors with transitive-only verbs (e.g. *The ball kicked) and
transitivisation errors with intransitive-only verbs (e.g. *Homer swam the fish).
Experiment 1 provided strong support for the semantics hypothesis, with semantic
factors significantly predicting difference scores for both transitive-only and
intransitive-only verbs. Overall verb frequency also significantly predicted difference
scores for both verb types, with difference scores increasing as verb frequency
increased. This is the pattern of results expected according to the entrenchment
hypothesis and thus provides strong support for this account. In contrast, the
preemption hypothesis was not supported in Experiment 1, with the only significant
predictive relationship being in the unexpected direction.
Results for Experiment 2 were analysed separately for each age group due to
significant by-age interactions in the initial analyses. For the youngest children,
nothing predicted difference scores for intransitivisation errors with transitive-only
verbs, although scores for transitivisation errors with intransitive-only verbs were
predicted by one semantic factor and total verb frequency. These predictions were
both in the expected direction, therefore providing some support for the semantics
and entrenchment hypotheses respectively. For the older children, total verb
frequency predicted difference scores for both verb types, thus providing further
strong support for the entrenchment hypothesis. Again, though, no support was found
for the preemption hypothesis. The role of semantics here is less clear, as semantic
factors were significant predictors of difference scores, but not in the expected
directions. Finally, for adults, neither entrenchment nor preemption found any
support. While various semantic factors significantly predicted difference scores for
adults, the direction of these predictions was often in the unexpected direction, just
like with the older children, meaning the role of semantics is unclear. The role of
entrenchment and semantics also seems to change over the age groups in Experiment
2, with entrenchment becoming less important as age increases, but semantics
potentially increasing in its influence.
Experiment 3 used production-priming in a novel way: as a method to elicit
overgeneralisation errors from 5- to 6-year-old children. While grammaticality
judgment studies tell us what participants consciously know about overgeneralisation
errors, these errors are primarily a phenomenon of production in young children.
161
Testing the predictions of our hypotheses against the errors that children actually
produce was therefore an important test, and something that is almost impossible to
do using data from corpora or diary studies, primarily due to their sparsity in
naturally-occurring data. We were successful in eliciting large numbers of both
intransitivisation errors with transitive-only verbs and transitivisation errors with
intransitive-only verbs. Semantic factors predicted error rates in both directions,
whereas the entrenchment and preemption predictors predicted error rates only for
transitivisation errors with intransitive-only verbs. In Experiment 3, then, support
was found for all three hypotheses under investigation.
Taken together, the experiments reported in Chapter 4 provide strong support
for the entrenchment hypothesis, with total verb frequency predicting difference
scores and error rates in the expected direction for many parts of the experiments.
Very little support for the preemption hypothesis was found, however, with the only
positive evidence coming from the production of transitivisation errors in
Experiment 3 (although this is perhaps the best test of this hypothesis). Evidence for
the semantic account was somewhat mixed: although semantic factors predicted
results in almost all parts of the experiments, some of the predictions were in the
opposite direction to our expectations. Again, evidence from Chapter 4 points to the
need for an integrated account that takes into consideration the observed effects of
entrenchment, preemption and semantics.
6.2.3 Chapter 5: Passives
Chapters 3 and 4 focussed on children’s retreat from argument structure
overgeneralisation errors, with these studies finding evidence for the influence of
both semantic and statistical factors. However, any mechanism involved in language
acquisition should also be able to explain effects outside of the limited domain of
overgeneralisations. This final experimental chapter, therefore, examined the
influence of semantics in children’s acquisition of the passive, a construction that is
notoriously difficult to master.
Taking the study of Messenger et al. (2012) as a starting point, this chapter
used structural priming to encourage children to produce passive sentences to
describe a series of animations. Previous studies (e.g. Maratsos et al., 1985) found
that children’s difficulties in comprehension of passive sentences varied by verb
162
type, with agent-patient (AP) verbs (e.g. hold) proving relatively easy to comprehend
in comparison with experiencer-theme (ET) verbs (e.g. see). While Messenger et al.
replicated these findings in a comprehension task, a priming task found that children
were primed equally by passive sentence containing verbs of three different semantic
classes – AP, ET and TE (theme-experiencer verbs, e.g. frighten) – to produce
passive sentences with AP verbs. This led the authors to conclude that children, in
fact, already have an abstract representation of the passive construction at the age of
three years, and that the differences between the semantic classes in comprehension
tasks were likely due to the relative difficulty in depicting ET events in still pictures.
In this chapter, we reversed Messenger et al.’s method, using only AP primes
but target verbs from all three semantic classes. We found priming effects for all verb
types, thus replicating Messenger et al.’s finding that young children already have an
abstract representation of the passive (as they would not be able to be primed without
having this), but we also found significant differences in rate of production of
passives between the verb classes. Both adults and children produced the greatest
proportion of passive sentences with TE verbs, a smaller proportion with AP verbs
and the smallest proportion with ET verbs. This result was replicated using a more
fine-grained approach to semantics, similar to that taken in Chapter 4 (see also
Ambridge, Pine, Rowland & Chang, 2012, and Ambridge et al., 2014). For this
analysis, we used a composite semantic factor derived via Principal Components
Analysis, based on adult judgments (see Ambridge, Bidgood, Pine, Rowland &
Freudenthal, 2015). This enabled us to treat semantics as a continuum, based on the
core semantics of the passive construction – that of Affectedness of the passive
subject. The results of this analysis showed that, the more a verb’s semantics fitted
with these semantics, the more likely both children and adults were to produce
passive sentences with that verb.
These findings led us to conclude that a semantic prototype account is the best
explanation for the acquisition of the passive: just as a robin is a more prototypical
member of the ‘bird’ category than an ostrich, TE verbs such as frighten, which are
high in Affectedness, fit better with the prototypical meaning of the passive
construction than ET verbs such as see, which are relatively low in Affectedness.
These findings are in line with the recent comprehension and grammaticality
judgment studies of Ambridge, Bidgood, Pine, Rowland and Freudenthal (2015). The
findings also add further support to the idea that the semantic fit between the verb
163
and the construction, as proposed by the FIT account (Ambridge & Lieven, 2011;
Ambridge et al., 2011), plays an important role in the selection of an appropriate
construction when conveying a message.
6.3 Summary
The studies presented in this thesis investigated three hypotheses: the semantic verb
class hypothesis (Pinker, 1989), the entrenchment hypothesis (Braine & Brooks,
1995) and the preemption hypothesis (Goldberg, 1995). Between the studies, support
was found for all three hypotheses, although evidence for the preemption hypothesis
was less convincing than for semantics and entrenchment. This section will
summarise the new evidence for these theories presented in the thesis. Finally, a
discussion of how this evidence supports the need for an integrated account, such as
the FIT account (Ambridge & Lieven, 2011; Ambridge et al., 2011), will be
discussed.
6.3.1 Evidence for the semantics hypothesis
The role of semantics was tested in all three chapters in this thesis. The initial
approach, in Chapter 3, was based on Pinker’s (1989) semantic verb class hypothesis.
This study was also a particularly stringent test of this hypothesis as it used novel
verbs: participants could not have used previous experience with these verbs in
context (as required for the operation of statistical-learning mechanisms) to provide
their grammaticality judgments and must have made these judgments on the basis of
semantic information. The fact that children and adults were able to able to provide
grammaticality judgments for locative sentences containing the novel verbs in line
with real verbs from the same semantic class, for at least two of the three verb classes
tested in each age group, provides strong support for the semantic verb class
hypothesis. These findings are in line with previous work (e.g. Ambridge et al.,
2008; 2009; 2011; Brooks & Tomasello, 1999; Gropen et al., 1991a, b).
Following more recent work by Ambridge and colleagues (e.g. Ambridge,
Pine & Rowland, 2012; Ambridge, Pine, Rowland & Chang, 2012; Ambridge et al.,
2014), Chapter 4 took a slightly different approach to the influence of semantics.
Rather than viewing semantics in terms of discrete verb classes, we created
164
continuous semantic variables, related to the causative alternation, by collecting
semantic ratings from adults and using Principal Components Analysis (PCA) to
reduce these to a smaller number of composite semantic factors. These were then
used as the semantic predictors in the statistical models. This approach allows more
flexibility to investigate widely-observed lexical effects. Results of the analysis
suggested that semantics did indeed have an influence on the way in which both
adults and children judge the grammaticality of overgeneralisation errors in the
causative alternation, and on the rate at which young children actually produce
overgeneralisation errors. However, the direction of prediction of each of the
composite semantic factors was not always as expected, and sometimes appeared to
be contradictory. One possible reason for this is that our labelling of the semantic
factors produced by the PCA was not reflective of the complex make-up of each of
these factors, since each was created from 26 original semantic features (see Chapter
4 for further discussion). Nevertheless, the fact remains that semantics did have an
effect both on grammaticality judgments and error production.
So far, then, semantics has been shown to have an influence on children’s
retreat from overgeneralisation errors. Chapter 5 set out to test if semantics also has a
role to play in a more general language acquisition mechanism. Since work with
adults (Ambridge, Bidgood, Pine, Rowland & Freudenthal, 2015) has demonstrated
an influence of semantics on adults’ comprehension and grammaticality judgments
of (primarily grammatical) passive sentences, and given that the passive is a
construction that children are known to struggle with, in comprehension at least, until
a relatively advanced age (7 years in Maratsos et al., 1985), Chapter 5 tested whether
children’s and adults’ production of the passive would similarly be affected by verb
semantics. As this study was based on the methodology of Messenger et al. (2012),
we tested semantics using the three semantic classes/types (AP, ET and TE) used in
her study, as well as using a semantic continuum (borrowed from Ambridge,
Bidgood, Pine, Rowland & Freudenthal, 2015). Although results support Messenger
et al.’s finding that even young children have an abstract representation of the
passive construction, a clear influence of semantics was also found, for both adults
and children, using both the class-based and fine-grained semantic continuum
approaches. The conclusion from this is that the abstract syntactic knowledge that is
already in place at 4 years of age is, nonetheless, semantically constrained. Thus,
165
semantics has a wider influence on language development beyond its role in
children’s retreat from overgeneralisation errors.
Investigating both class-based and fine-grained continuum approaches to
semantics in the same study allowed for comparison of these two approaches. Results
suggest that the semantic continuum approach is the best fit for the data: Figure 5.3
showed that verbs in all three of the original classes (AP, TE and ET) vary in their
level of Affectedness, and that the proportion of passive sentences produced by
participants varied in line with this. In addition, the level of Affectedness (the
prototypical meaning associated with the passive construction) for verbs in each of
the classes overlapped with one other. In summary, while elements of the class-based
semantics approach proposed by Pinker (1989) have merit (the fine-grained semantic
factors used in both Chapters 4 and 5 were derived from these, after all), evidence
points to the need to take a more fine-grained, graded approach to semantics, in line
with that proposed by the FIT account (Ambridge & Lieven, 2011; Ambridge et al.,
2011). (Note that the use of a Likert scale, rather than categorical
grammatical/ungrammatical judgments in Experiments 1 and 2 in Chapter 4 may
have forced the participants into providing judgments that favour a continuous scale.
Pinker’s original [1989] semantic verb class hypothesis is therefore not fully
countered by the findings of this thesis.)
6.3.2 Evidence for the entrenchment hypothesis
The entrenchment hypothesis was tested in Chapters 3 and 4 (although it was not
distinguished from preemption in Chapter 3, and was labelled the frequency
hypothesis). In Chapter 3’s investigation of locative overgeneralisation errors, verb
frequency was treated as a categorical variable, with verbs classed as high-frequency,
low-frequency or novel (essentially zero-frequency), based on corpus data. Results
showed strong support for the frequency hypothesis: participants showed a
significantly larger dispreference for overgeneralised sentences (compared to their
grammatical counterparts) containing high-frequency verbs than either low-
frequency or novel verbs, as well as a larger dispreference for overgeneralised
sentences containing low-frequency verbs than those containing novel verbs.
In Chapter 4, entrenchment was the most consistent of our predictors across
the grammaticality judgment and production studies. In Experiment 1, adults’
166
grammaticality judgments were predicted by our measure of entrenchment: total verb
frequency (determined by corpus counts). In Experiment 2, total verb frequency
predicted grammaticality judgments for both groups of children tested, although this
was no longer the case for adults. This discrepancy from Experiment 1 could have
been related to the smaller set of verbs used, which were mostly the higher-frequency
verbs used in Experiment 1. This result may indicate a developmental effect of
entrenchment, and may also be indicative of a ceiling effect: perhaps it is not
possible for verbs to become yet more entrenched once they have been experienced a
certain (presumably large) number of times. Importantly, verb frequency also
predicted the rate at which children produced overgeneralisation errors with
individual verbs: the higher the frequency of the verb, the less likely children were to
produce an overgeneralisation error with that verb.
Taken together, then, the studies in Chapters 3 and 4 provide strong support
for the role of entrenchment in children’s retreat from overgeneralisation errors, in
line with previous work (e.g. Ambridge et al., 2008, 2011; Brooks et al., 1999;
Theakston, 2004). Chapter 4 also raises interesting possibilities for future research
related to developmental effects of entrenchment, and a possible ceiling effect.
6.3.3 Evidence for the preemption hypothesis
Like the entrenchment hypothesis, the preemption hypothesis was tested in Chapters
3 and 4. However, in Chapter 3, the high correlation between total verb frequency
and verb frequency in the preempting construction (here, the alternative locative
construction) made it infeasible to distinguish between the predictions of preemption
and entrenchment (see also Ambridge, Pine & Rowland, 2012; Boyd et al., 2012;
Wonnacott, 2011, p. 2; Perfors et al., 2010, p. 612). Thus, the findings in support of
the entrenchment hypothesis in Chapter 3, described above, apply equally to the
preemption hypothesis.
Chapter 4 tested the preemption hypothesis directly. Following Brooks and
Tomasello (1999) and Brooks and Zizak (2002), we used the passive as the
preemptive construction for avoiding intransitivisation errors with transitive-only
verbs (e.g. The ball was kicked preempts *The ball kicked) and the periphrastic
causative as the preemptive construction for avoiding transitivisation errors with
intransitive-only verbs (e.g. Homer made the fish swim preempts *Homer swam the
167
fish). The results of the experiments provided somewhat mixed evidence for the
preemption hypothesis. In Experiment 1, preemption predicted difference scores for
adults’ grammaticality judgments of intransitivisation errors with transitive-only
verbs, but in the opposite direction to our expectations: the more a transitive-only
verb had appeared in the passive, the more likely adults were to accept it in an
overgeneralised intransitive sentence. This contradicts the preemption hypothesis. In
Experiment 2, preemption did not predict grammaticality judgments at all. Thus,
these studies found no evidence that preemption plays a role in grammatical
acceptability judgments of intransitivisation errors with transitive-only verbs or
transitivisation errors with intransitive-only verbs. This goes against the findings of a
previous study using the same methodology with the dative alternation (Ambridge et
al., 2014) and suggests that preemption might play a role for the retreat from
overgeneralisation errors in some alternations but not others (see also Ambridge,
Pine & Rowland’s 2012 grammaticality judgment study of the locative alternation, in
which no dissociable effect was found between entrenchment and preemption).
Chapter 4’s Experiment 3, in the other hand, did find a significant effect of
preemption. The rate at which children produced transitivisation errors with
intransitive-only verbs was significantly predicted by verb frequency in the
preempting periphrastic causative construction (although intransitivisation errors
with transitive-only verbs were not predicted by our preemption measure). The task
effect seen here (see also Blything et al., 2014) might be due to competition between
constructions: when producing a sentence to convey the desired message, all possible
constructions are competing to express the message, thus forcing a choice between
the alternative constructions. In a grammaticality judgment task, however, this is not
necessarily the case. The results of Experiment 3, then, do lend support to the
preemption hypothesis, suggesting that it may play a role in the production of
overgeneralisation errors which is, after all, the phenomenon it is designed to
explain.
Taken together, these findings suggest that the role preemption plays in
children’s retreat from overgeneralisation errors is not as clear-cut as those played by
semantics and entrenchment. That role may differ by construction and might depend
on the demands of the task at hand, with production tasks perhaps providing the best
evidence that a preemption mechanism does, indeed, have a role to play.
168
6.3.4 Comparing frequency effects across studies
Verb frequency measures were in included in the analyses in Chapters 3 and 4. In
Chapter 3, a single frequency measure was included as a categorical variable: verbs
were classified as high or low frequency (or novel). Frequency effects were observed
as expected, with all participants judging ungrammatical sentences to be less
acceptable with high-frequency than low-frequency verbs. In Chapter 4, total verb
frequency counts were used as the entrenchment predictor with verb frequency in a
competing construction with similar semantics as the preemption predictor. Results
in this study were not straightforward. Firstly, neither entrenchment nor preemption
were significant predictors in all studies. As discussed above, this could be due to
these mechanisms operating differently with the different ages (Experiment 2
appeared to suggest a ceiling effect for entrenchment) and different tasks
(preemption seemed to be a better predictor in the production task). Furthermore, the
unexpected direction of the prediction of our preemption predictor in Experiment 1
could have been influenced by the large number of additional predictors in the
model: a preemption-only model might have given different results. In addition, it is
worth noting that the frequency measures seemed to be more successful predictors
for transitivisation errors with intransitive-only verbs (e.g. *Homer swam the fish)
than intransitivisation errors with transitive-only verbs (e.g. *The ball kicked). These
apparent construction-specific effects are not entirely surprising, however. Previous
studies by Ambridge and colleagues have suggested that frequency information may
be used differently in locative and dative construction pairs, for example (Ambridge,
Pine & Rowland [2012] and Ambridge et al. [2014], respectively). Nevertheless, it
potentially dangerous to draw firm conclusions about the frequency effects in these
studies (particularly the apparent null effects) due to the small numbers of
participants in the studies (see Tversky & Kahneman, [1971] and Dienes [2014]).
Overall, frequency effects were found, to some extent, in all studies in which
they were examined. In fact, even though Chapter 5 did not include a frequency
measure, apparent frequency effects were suggested by the analysis presented in
Figure 5.3. As this figure show, a group of theme-experiencer verbs (e.g. please,
worry, amaze) were produced by children far more frequently in the passive
construction than other verbs of that class. It is likely that these verbs are heard
frequently in truncated passive/adjectival uses (e.g. I was amazed), which might have
169
increased the likelihood with which children were likely to produce them in full
passives here. Frequency effects in the production of passives would be in line with
the comprehension/reaction-time and grammaticality judgment studies (with adults)
of Ambridge, Bidgood, Pine, Rowland and Freudenthal (2015) and Street and
Dąbrowska (2014).
6.3.5 The FIT account: An integrated approach
The evidence presented above suggests that children’s retreat from
overgeneralisation errors is influenced by verb-in-construction semantics, verb
frequency in the input and, perhaps to a lesser extent, the frequency of those verbs in
particular (preempting) constructions. None of these factors can explain all of the
variance in adults’ and children’s responses to grammaticality judgment tasks, nor
their performance in production-priming tasks. Several accounts have been proposed
that integrate aspects of semantics and statistics to explain aspects of children’s
language acquisition, including, including Langacker (2000), MacWhinney (2004)
and Tomasello (2003). The FIT account, proposed and developed by Ambridge and
colleagues (e.g. Ambridge & Lieven, 2011; Ambridge, Pine & Rowland, 2011), is a
more recent version of these accounts, incorporating aspects of each of them. By
doing so, it aims to provide a more complete picture of children’s retreat from
overgeneralisation errors, and of their language development more generally. It is
worth noting that the FIT account is still being developed, with several aspects still
underspecified. For example, the precise way in which the factors discussed below
(verb frequency, construction frequency, etc.) interact with each other is not well-
defined, nor, in fact, are some of the factors themselves. This lack of specificity
means that the predictions of the FIT account are not necessarily clear. Nevertheless,
to the extent that is currently possible, this section will discuss the FIT account in the
context of the findings of the current thesis.
The FIT account proposes that the effects of semantics, entrenchment and
preemption can all be explained by constructions competing to convey the speaker’s
message. If the message the speaker wants to convey is that a particularly amusing
joke made a girl laugh, initially, all constructions will be available for the speaker to
use. Constructions that appear more frequently in the language environment, such as
the transitive-causative and intransitive-inchoative constructions, are initially likely
170
to be more highly activated. (Recently-heard constructions will also have increased
levels of activation, as indicated by structural priming effects.) The relevance of the
competing constructions to convey the entire message increases the activation of
some constructions relative to others. In this case, the most relevant constructions are
those with slots for the verb, laugh, and each of the participants, joke and girl, such
as the transitive-causative and the periphrastic causative. Constructions that are
completely irrelevant, particularly those of low frequency (e.g. the locative
constructions), may be dismissed. The frequency of the verb in each of the
constructions also plays a role – as laugh has been frequently heard in the
intransitive-inchoative construction, its activation level increase relative to the
transitive-causative, despite it having less relevance for conveying the message.
Finally, and importantly, the semantic compatibility (or ‘fit’) between the
[ACTION] slot in the competing constructions and the semantics of the verb come
into play. The [ACTION] slot in the transitive-causative construction implies direct,
physical causation. While causation between the joke and the girl laughing does
exist, it is neither direct nor physical – the joke amused the girl and the amusement
caused the physical act of laughing. Thus, the semantics of laugh do not fit with the
semantics of the transitive-causative construction. The semantics of the [ACTION]
slot in the intransitive-inchoative construction, on the other hand, relate to internally-
caused events. While this ignores the causal aspect of the message (and makes the
intransitive-inchoative construction less relevant to conveying the entire message),
the verb is still compatible with the construction – the girl’s internal amusement at
the joke caused the physical response. The [ACTION] slot in the periphrastic
causative implies indirect causation. This [ACTION] slot is therefore an excellent
semantic fit for the verb, as well as the construction itself being highly relevant for
conveying the message. However, the periphrastic causative is disadvantaged in
terms of both its overall frequency in the input and, in comparison with the
intransitive-inchoative construction, the verb-in-construction frequency.
Thus, with relevance to the message, overall frequency, verb-in-construction
frequency and semantic compatibility all playing roles in the choice of construction,
the construction that will win out is not necessarily obvious. The construction with
the best fit, and possibly the most relevance for conveying the message, is probably
the periphrastic causative (The joke made the girl laugh). However, the verb’s
frequent appearance in the intransitive-inchoative construction, and its not-
171
incompatible semantics, may lead the intransitive-inchoative to be used (The girl
laughed). A child might produce the overgeneralised transitive-causative sentence,
*The joke laughed the girl, if the overall frequency of the construction itself (much
higher than the periphrastic causative) and its relevance to conveying the message (it
is more relevant than the intransitive-inchoative) outweigh the fact that the verb has
not been heard in the construction and that its semantic fit is not particularly good.
The FIT account clearly yields effects of semantics (through the compatibility
between the verb and the construction) and entrenchment (via verb-in-construction
frequency statistics). Preemption effects can also be seen, through a combination of
verb-in-construction frequency and verb-in-construction compatibility. Importantly,
the differences in the efficacy of preemption to explain the retreat from
overgeneralisation in different construction pairs can also be explained via the
competition model. The two dative constructions (e.g. Lisa gave the book to
Bart/Lisa gave Bart the book) have high degrees of overlap in terms of their
semantics, their relative frequency in the input and the messages that they are likely
to be relevant for conveying. Thus, the frequency with which a verb appears in one
or other construction in the input will have a large effect on how likely it is to be
chosen over the other. A clear effect of preemption will therefore be seen, as in
Ambridge et al. (2014). However, when the competing constructions are very rare in
the input, as with the locative construction, the preemption effect may be too small to
be seen over and above a more general entrenchment effect (see Ambridge, Pine &
Rowland, 2012).
In contrast to the dative alternation, the preempting structures for
overgeneralisations errors in the transitive-intransitive alternation, investigated in
Chapter 4 of this thesis, are generally considered to be different constructions
entirely (e.g. Brooks & Tomasello, 1999) – the passive construction (e.g. The ball
was kicked) preempts intransitivisation errors with transitive-only verbs (e.g. *The
ball kicked), whereas the periphrastic causative construction (e.g. Homer made the
fish swim) preempts transitivisation errors with intransitive-only verbs (e.g. *Homer
swam the fish). In addition, both of these preempting constructions are much lower in
frequency than the transitive and the intransitive constructions themselves. This
makes preemption effects much more difficult to observe, and goes some way to
explaining the finding that no preemption effects were apparent in the judgment
studies in Chapter 4.
172
The competition mechanism is clearly effective in explaining both children’s
initial production of overgeneralisation errors and how they might retreat from these.
The mechanism has recently been instantiated as a connectionist model (Ambridge &
Blything, in press) using the dative alternation (e.g. Paul gave the book to Mary/Paul
gave Mary the book), with the model producing the pattern of overgeneralisation and
retreat from overgeneralisation observed in children’s acquisition data. This model
was also able to correctly predict the construction(s) in which novel verbs would be
grammatical using semantics alone (as adults and children were able to do in Chapter
3 of this thesis) and to reproduce the pattern of grammaticality judgments given by
adults in Ambridge, Pine, Rowland and Chang (2012). These findings therefore offer
strong support for the FIT account as a learning model that can truly account for the
data observed in both production and grammaticality judgment studies.
This thesis has provided additional evidence for the FIT account, and thus
enabled further development of thoughts about its operation, in a number of ways.
The locatives study in Chapter 3 demonstrated that both semantic and frequency
elements of the FIT account are in operation with this construction pair. While
Ambridge, Pine and Rowland (2012) found similar results, the current study had the
advantage that it also used novel verbs as a conclusive test of the role of verb
semantics, showing that children and adults are both able to use the semantic
compatibility between the new verb and the constructions alone (with no verb-in-
construction information at all) to make judgments in line with real verbs that have
similar semantics.
The causatives study in Chapter 4 showed that overall verb frequency
information, as well as semantics, predicted grammaticality judgments from both
adults and children. Verb frequency was used as the predictor variable to test the
entrenchment hypothesis, but it is also equivalent to the sum of all verb-in-
construction frequency counts, one of the four elements of the FIT account. Chapter
4 also presented some evidence for the use of semantic information (verb-in-
construction compatibility) in grammaticality judgments with the causative
alternation, although the nature of the continuous predictor variables created via
Principal Components Analysis meant that the exact nature of the relationship
between verb semantics and grammatical acceptability was unclear. In the production
study, effects of verb frequency (equivalent to the sum of verb-in-construction
frequencies), verb frequency in a competing construction (a combination of verb-in-
173
construction frequency and verb-in-construction compatibility) and semantics (verb-
in-construction compatibility) were observed. The fact that all of these effects were
significant in the same model adds weight to the argument that the various factors are
all working alongside each other to influence the eventual choice of competing
constructions.
Chapter 5 moved beyond overgeneralisation errors to look at the acquisition
of a construction known to be problematic for young children: the passive. This
chapter tested only the effect of verb-in-construction compatibility. The fact that both
children and adults were primed to produce more passive sentences with a verb the
more its semantics fit with those of the construction demonstrate that, not only does
the FIT account provide a viable mechanism for explaining the retreat from
overgeneralisation, but that this mechanism is able to apply more widely in the
acquisition of argument structure.
6.4 Possible future research directions
The research presented in this thesis has demonstrated that both semantic and
statistical accounts have a role to play in children’s retreat from overgeneralisation
errors and, in the case of semantics at least, in language acquisition more generally.
However, the findings of both Chapter 3 and, in particular, Chapter 4 emphasise that
further work is required to investigate how these mechanisms interact to enable
children to retreat from overgeneralisation errors or, indeed, avoid them altogether.
Since overgeneralisation errors are really a phenomenon of language production, the
priming method used in Chapters 4 and 5 seems a promising way to move forward,
rather than relying on metalinguistic knowledge using methodologies such as
grammaticality judgments (although this method has proven instrumental in
improving our understanding of the mechanisms in question up to this point). The
priming method could easily be extended to study dative and locative
overgeneralisation errors, for example.
Additionally, no mechanism for language acquisition can be language-
specific: children learn the language of their environment, irrespective of the
language of their biological parents. Nevertheless, the operation of this mechanism
might be somewhat different in languages other than English. One way to test this
would be to investigate how both statistical (entrenchment-like and preemption-like)
174
and semantic mechanisms might work together in languages that are typologically
different to English, but in which argument structure overgeneralisation errors are
still observed. For example, in some languages (e.g. K’iche’ Mayan), no systematic
variable in the morphological marker for the causative have been noted. Thus, effects
of statistical variable may be observed, while no effect of semantics would be
predicted. Hindi also has different morphological markers for causativisation, but
semantic differences have been noted in this language. A semantic effect would
therefore be expected. Conversely, while an effect of preemption would be predicted
(both morphological markers are effectively synonymous), no entrenchment effect
over and above this would be expected for Hindi due to the high frequency of the
two main competing constructions and the comparatively very low frequency of any
other constructions that can contain the verbs in question. By testing languages in
which not all mechanisms would seem feasibly able to operate, language-specific
and language-general aspects of the mechanism of argument structure acquisition can
be investigated.
A further avenue that ought to be explored is how these mechanisms relate to
language acquisition more generally, beyond the retreat from overgeneralisation.
Chapter 5 has begun to explore this question by investigating the role of semantics in
the production of passive sentences. This work could be extended to look at the roles
of semantics and entrenchment in children’s comprehension of the passive, following
Ambridge, Bidgood, Pine, Rowland and Freudenthal’s (2015) passive
comprehension work with adults. Other constructions should also be investigated,
partly because a successful mechanism involved in language acquisition cannot be
specific to a single construction (this would be too specific) but also because the
work presented here has suggested that the relative importance of different
mechanisms may we weighted differently with different argument structures.
Importantly, future studies should focus on how to test integrated accounts,
such as the FIT account, rather than continuing to test entrenchment, preemption and
semantic mechanisms separately.
6.5 Conclusion
The three experimental chapters of this thesis have demonstrated clear effects of both
frequency and semantics in children’s retreat from overgeneralisation errors. They
175
have replicated and extended the results of previous papers in this area through the
use of improved and innovative methodologies. While Chapter 4 provided clear
evidence for the entrenchment hypothesis, however, no strong evidence for the
preemption hypothesis was found. In addition, Chapter 5 demonstrated that semantic
effects are to be found in other domains of language acquisition; here, in the
acquisition of the passive construction. Despite the evidence for each of the
hypotheses presented here, what is clear is that none of them can explain all of the
lexical effects observed in the data on its own. What is needed is an approach that
integrates both semantic and frequency information into account in its explanation of
children’s retreat from error, such as the FIT account (Ambridge & Lieven, 2011;
Ambridge et al., 2011). While the current paper has provided important additional
evidence for this account, it has also demonstrated that much more work is needed to
clarify how the factors of construction frequency, verb-in-construction frequency,
verb-in-construction compatibility and relevance interact with each other in the
competition between constructions to convey a speaker’s message.
177
References
Abbot-Smith, K., Lieven, E., & Tomasello, M. (2001). What preschool children do
and do not do with ungrammatical word orders. Cognitive Development,
16(2), 679-692.
Agresti, A. (2002). Categorical data analysis Vol. 359. Hoboken, NJ: John Wiley &
Sons.
Akhtar, N. (1999). Acquiring basic word order: Evidence for data-driven learning of
syntactic structure. Journal of Child Language, 26(2), 339-356.
Alishahi, A. & Stevenson, S. (2008). A computational model of early argument
structure acquisition. Cognitive Science, 32(5), 789-834.
Altman, G. T. M., Dienes, Z., & Goode, A. (1995). Modality Independence of
Implicitly Learned Grammatical Knowledge. Journal of Experimental
Psychology: Learning, Memory and Cognition, 21(4), 899-912.
Ambridge, B. (2011). Paradigms for assessing children’s knowledge of syntax and
morphology. In Hoff, E. (Ed.). Guide to Research Methods in Child
Language. London: Blackwell-Wiley, pp. 113-132.
Ambridge, B. (2013). How do children restrict their linguistic generalizations?: An
(un-)grammaticality judgment study. Cognitive Science, 37(3), 508-543.
Ambridge, B., Bidgood, A., Pine, J. M., Rowland, C. F., & Freudenthal, D. (2015). Is
passive syntax semantically constrained? Evidence from adult grammaticality
judgment and comprehension studies. Cognitive Science.
Ambridge, B., Bidgood, A., Twomey, K. E., Pine, J. M., Rowland, C. F., &
Freudenthal, D. (2015). Preemption versus entrenchment: Towards a
construction-general solution to the problem of the retreat from verb
argument structure overgeneralisation. PLoS ONE, 10(4): e0123723.
Ambridge, B., & Blything, R. P. (in press). A connectionist model of the retreat from
verb argument structure overgeneralization. Journal of Child Language.
Ambridge, B. & Lieven, E. V. M. (2011). Child Language Acquisition: Contrasting
theoretical approaches. Cambridge: Cambridge University Press.
Ambridge, B. & Lieven, E. V. M. (2015). A constructivist account of child language
acquisition. In B. MacWhinney & W. O’Grady (Eds.), Handbook of
Language Emergence (pp. 478-510). Hoboken, NJ: Wiley Blackwell.
178
Ambridge, B., Pine, J. M. & Rowland, C. F. (2011). Children use verb semantics to
retreat from overgeneralization errors: A novel verb grammaticality judgment
study. Cognitive Linguistics, 22(2), 303-323.
Ambridge, B., Pine, J. M. & Rowland, C. F. (2012). Semantics versus statistics in the
retreat from locative overgeneralization errors. Cognition, 123, 260-279.
Ambridge, B., Pine, J. M., Rowland, C. F. & Chang, F. (2012). The roles of verb
semantics, entrenchment, and morphophonology in the retreat from dative
argument-structure overgeneralization errors. Language, 88(1), 45-81.
Ambridge, B., Pine, J. M, Rowland, C. F., Freudenthal, D., & Chang, F.
(2014). Avoiding dative overgeneralization errors: Semantics, statistics or
both? Language, Cognition and Neuroscience, 29(2), 218-243.
Ambridge, B., Pine, J. M., Rowland, C. F., Jones, R. L. & Clark, V. (2009). A
semantics-based approach to the "no negative evidence" problem. Cognitive
Science, 33(7), 1301-1316.
Ambridge, B., Pine, J. M., Rowland, C. F. & Young, C. R. (2008). The effect of verb
semantic class and verb frequency (entrenchment) on children's and adults'
graded judgements of argument-structure overgeneralization errors.
Cognition, 106(1), 87-129.
Anderson, S. R. (1971). On the role of deep structure in semantic interpretation.
Foundations of Language, 7, 387-396.
Anime Studio Pro Version 5.5 (2006). Scotts Valley, CA: e frontier America, Inc.
Baker, C. L. (1979). Syntactic theory and the projection problem. Linguistic Enquiry,
10, 533-581.
Baker, M. C. (2003). Lexical categories: Verbs, nouns and adjectives (Cambridge
Studies in Linguistics, vol. 102). Cambridge: Cambridge University Press.
Bannard, C. & Matthews, D. (2008). Stored word sequences in language learning.
Psychological Science, 19(3), 241-248.
Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure
for confirmatory hypothesis testing: Keep it maximal. Journal of memory and
language, 68(3), 255-278.
Bates, D., Maechler, M., Bolker, B., & Walker, S. (2014). lme4: Linear mixed-
effects models using Eigen and s4. R package version 1.1-7. http://cran.r-
project.org/package=lme4.
179
Bencini, G. M. L., & Valian, V. V. (2008). Abstract sentence representations in 3-
year-olds: Evidence from language production and comprehension. Journal
of Memory and Language, 59(1), 97-113.
Bernaards, C. A. & Jennrich, R. I. (2005). Gradient projection algorithms and
software for arbitrary rotation criteria in factor analysis. Education and
Psychological Measurement: 65, 676-696.
http://www.stat.ucla.edu/research/gpa.
Bidgood, A., Ambridge, B., Pine, J. M., & Rowland, C. F. (2014). The retreat from
locative overgeneralisation errors: A novel verb grammaticality judgment
study. PLoS ONE, 9(5), e97634.
Bidgood, A., Ambridge, B., Pine, J. M. & Rowland, C. F. (in prep.). Children’s and
adults’ passive syntax is semantically constrained: Evidence from syntactic
priming.
Bidgood, A., Ambridge, B., Pine, J. M., Rowland, C. F., & Freudenthal, D. (in
prep.). How do children retreat from overgeneralisation error? Evidence from
the causative alternation.
Blything, R., Ambridge, B., & Lieven, E. V. M. (2014). Children use statistics and
semantics in the retreat from overgeneralization. PLoS ONE, 9(10) e110009.
Boas, H. C. (2008). Towards a frame-constructional approach to verb classification.
Revista Canaria de Estudios Ingleses, 57, 17-48.
Bock, J. K. (1986). Syntactic persistence in language production. Cognitive
Psychology, 18(3), 355-387.
Bock, K., & Loebell, H. (1990). Framing sentences. Cognition, 35(1), 1-39.
Bock, K., Loebell, H., & Morey, R. (1992). From conceptual roles to structural
relations: Bridging the syntactic cleft. Psychological Review, 99(1), 150-171.
Borer, H., & Wexler, K. (1987). The maturation of syntax. In T. Roeper & E.
Williams (Eds.), Parameter setting (pp. 123-172). Dordrecht: D. Reidel
Publishing Company.
Bowerman, M. (1973). Early syntactic development: A cross-linguistic study with
special reference to Finnish. Cambridge: Cambridge University Press.
Bowerman, M. (1976). Semantic factors in the acquisition of rules for word use and
sentence construction. In Directions in normal and deficient language
development (pp. 99-179). Baltimore: University Park Press.
180
Bowerman, M. (1978). Systematising semantic knowledge: changes overtime in the
child’s organisation of word meaning. Child Development, 49, 977-987.
Bowerman, M. (1981). The child’s expression of meaning: Expanding relationships
among lexicon, syntax and morphology. Annuls of the New York Academy of
Sciences 379(1), 172-189.
Bowerman, M. (1982a). Evaluating competing linguistic models with language
acquisition data: Implications of developmental errors with causative verbs.
Quaderni di Semantica, 3(1), 5-66.
Bowerman, M. (1982b). Reorganizational processes in lexical and syntactic
development. In Wanner, E. & Gleitman, L. R. (Eds.), Language acquisition:
the state of the art. Cambridge: Cambridge University Press, pp. 319-346.
Bowerman, M. (1988). The "no negative evidence" problem: how do children avoid
constructing an overly general grammar? In Hawkins, J. A. (Ed.), Explaining
language universals. Oxford: Blackwell, pp. 73-101.
Bowerman, M. (1996). Argument structure and learnability: Is a solution in sight?
Proceedings of the 22nd Annual Meeting of the Berkley Linguistics Society:
General Session and Parasession on The Role of Learnability in
Grammatical Theory, 454-468.
Boyd, J. K. & Goldberg, A. E. (2011). Learning what NOT to say: The role of
statistical preemption and categorization in A-adjective production.
Language, 87(1), 55-83.
Boyd, J. K., Ackerman, F. & Kutas, M. (2012). Adult learners use both entrenchment
and preemption to inter grammatical constraints. In 2012 IEEE International
Conference on Development and Learning and Epigenetic Robots. Red Hook,
NY: Curran Associates, pp. 1-2.
Braine, M. D. S. (1963). The ontogeny of English phrase structure: The first phase.
Language, 39(1), 1-13.
Braine, M. D. S. (1971). On two types of models of the internalization of grammars.
In D. I. Slobin (Ed.), The Ontogenesis of Grammar (pp. 153-186). New York:
Academic Press.
Braine, M. D. S. (1976). Children’s first word combinations. Monographs of the
Society for Research in Child Development, 41(164), 1-104.
181
Braine, M. D. S., Brody, R. E., Fisch, S. M., Weisberger, M. J., & Blum, M. (1990).
Can children use a verb without exposure to its argument structure? Journal
of Child Language, 17(2), 313-342.
Braine, M. D. S. & Brooks, P. J. (1995). Verb argument structure and the problem of
avoiding an overgeneral grammar. In Tomasello, M. & Merriman, W. E.
(Eds.), Beyond names for things: Young children's acquisition of verbs.
Hillsdale, NJ: Erlbaum, pp. 352-376.
Branigan, H. (2007). Syntactic priming. Language and Linguistics Compass, 1(1-2),
1-16.
Branigan, H., Pickering, M. J., Liversedge, S. P., Stewart, A. J., & Urbach, T. P.
(1995). Syntactic priming: Investigating the mental representation of
language. Journal of Psycholinguistic Research, 24(6), 489-506.
Bresnan, J. (2001). Lexical-Functional Syntax. Oxford: Blackwell.
Bresnan, J., Cueni, A., Nikitina, T., & Baayen, R. H. (2007). Predicting the dative
alternation. Cognitive Foundations of Interpretation, 69-94.
Brinkmann, U. (1997). The locative alternation in German: Its structure and
acquisition. Amsterdam: John Benjamins Publishing Co.
British National Corpus (2007). BNC XML edition (3rd Ed..). Oxford: Oxford
University Computing Services (distributor), on behalf of the BNC
Consortium.
Brooks, P. J. & Tomasello, M. (1999). How children constrain their argument
structure constructions. Language, 75(4), 720-738.
Brooks, P. J., Tomasello, M., Dodson, K. & Lewis. L. B. (1999). Young children's
overgeneralizations with fixed transitivity verbs. Child Development, 70(6),
1325-1337.
Brooks, P. J. & Zizak, O. (2002). Does preemption help children learn verb
transitivity? Journal of Child Language, 29, 759-781.
Brown, R. (1973). A first language: The early stages. Cambridge, MA: Harvard
University Press.
Brown R., & Hanlon, C. (1970). Derivational complexity and order of acquisition in
child speech. In J. R. Hayes (Ed.), Cognition and the development of
language (pp. 11-53). New York: Wiley.
182
Cameron-Faulkner, T., Lieven, E., & Theakston, A. (2007). What part of no do
children not understand? A usage-based account of multiword negation.
Journal of Child Language, 34(2), 251-282.
Chan, A., Meints, K., Lieven, E., & Tomasello, M. (2010). Young children’s
comprehension of English SVO word order revisited: Testing the same
children in act-out and intermodal preferential looking tasks. Cognitive
Development, 25(1), 30-45.
Chang, F., Dell, G. S. & Bock, K. (2006). Becoming syntactic. Psychological
Review, 113(2), 234-272.
Childers, J. B., & Tomasello, M. (2002). The role of pronouns in young children’s
acquisition of the English transitive construction. Developmental Psychology,
37(6), 739-748.
Chomsky, N. (1980). Rules and representations. Oxford: Basil Blackwell.
Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris.
Chomsky, N. (1993). A minimalist program for linguistic theory. MIT occasional
papers in linguistics no. 1. Cambridge, MA: MIT Working Papers in
Linguistics.
Chouinard, M. M. & Clark, E. V. (2003). Adult reformulations of child errors as
negative evidence. Journal of Child Language, 30(3), 637-669.
Clark, E. (1988). On the logic of contrast. Journal of Child Language, 15(2), 317-
335.
Coyle, J. M., & Kaschak, M. P. (2008). Patterns of experience with verbs affect long-
term cumulative structural priming. Psychonomic Bulletin & Review, 15(5),
867–970.
Crain, S., & Nakayama, M. (1987). Structure dependence in grammar formation.
Language, 63(3), 522-543.
Crain, S., & Thornton, R. (2012). Syntax acquisition. Wiley Interdisciplinary
Reviews: Cognitive Science, 3(2), 185-203.
Dąbrowska, E., & Lieven, E. (2005). Towards a lexically specific grammar of
children’ question constructions. Cognitive Linguistics, 16(3), 437-474.
Dąbrowska, E., Rowland, C., & Theakston, A. (2009). The acquisition of questions
with long distance dependencies. Cognitive Linguistics, 20(3), 571-197.
183
Dąbrowska, E., & Street, J. (2006). Individual differences in language attainment:
Comprehension of passive sentences by native and non-native English
speakers. Language Sciences, 28(6), 604-615.
Demetras, M. J., Nolan Post, K., & Snow, C. E. (1986). Feedback to first language
learners: the role of repetitions and clarification questions. Journal of Child
Language, 13(2), 275-292.
Dienes, Z. (2014). Using Bayes to get the most out of non-significant results.
Frontiers in Psychology, 5, 781.
Dixon, R. M. W. (1979). Ergativity. Language, 55(1), 59-138.
Ezeizabarrena, M. J, & Larrañaga, M. P. (1996). Ergativity in Basque: A problem for
language acquisition? Linguistics, 34(5), 955-991.
Fellbaum, C. (1990). English verbs as a semantic net. International Journal of
Lexicography 3, 278-301.
Field, A., Miles, J., & Field, Z. (2012). Discovering statistics using R. London:
SAGE.
Fillmore, C. J. (1967). The case for case. In Bach, E. & Harms, R. (Eds.),
Proceedings of the Texas Symposium on Language Universals, April 13-15
1967. New York: Holt, Rinehart and Winston.
Fisher, C. (1996). Structural limits on verb mapping: The role of analogy in
children’s interpretations of sentences. Cognitive Psychology, 31(1), 41-81.
Fisher, C. (2002). The role of abstract syntactic knowledge in language acquisition:
A reply to Tomasello (2000). Cognition, 82(3), 259-278.
Fox, D., & Grodzinsky, Y. (1998). Children’s passive: A view from the by-phrase.
Linguistic Inquiry, 33(2), 311-332.
Gertner, Y., Fisher, C., & Eisengart, J. (2006). Learning words and rules: Abstract
knowledge of word order in early sentence comprehension. Psychological
Science, 17(8), 684-691.
Goldberg, A. E. (1995). Constructions: A construction grammar approach to
argument structure. Chicago: University of Chicago Press.
Goldberg, A. E. (2002). Surface generalizations: An alternative to alternations.
Cognitive Linguistics, 13(4), 327-356.
Goldberg, A. E. (2006). Constructions at work: The nature of generalization in
language. Oxford: Oxford University Press.
184
Goldberg, A. E. (2011). Corpus evidence of the viability of statistical preemption.
Cognitive Linguistics, 22, 131-154.
Gómez, R. L., & Gerken, L. (1999). Artificial grammar learning by 1-year-olds leads
to specific and abstract knowledge. Cognition, 70(2), 109-135.
Gómez, R. L., & Gerken, L. (2000). Infant artificial language learning and language
acquisition. Trends in Cognitive Science, 4(5), 178-186.
Gordon, P. & Chaftez, J. (1990). Verb-based versus class-based accounts of
actionality effects in children’s comprehension of passives. Cognition, 36(3),
227-254.
Gropen, J., Pinker, S., Hollander, M. & Goldberg, R. (1991a). Affectedness and
Direct Objects - the Role of Lexical Semantics in the Acquisition of Verb
Argument Structure. Cognition, 41(1-3), 153-195.
Gropen, J., Pinker, S., Hollander, M. & Goldberg, R. (1991b). Syntax and Semantics
in the Acquisition of Locative Verbs. Journal of Child Language, 18(1), 115-
151.
Gropen, J., Pinker, S., Hollander, M., Goldberg, R. & Wilson, R. (1989).
Learnability and acquisition of the dative alternation in English. Language,
62(2), 203-257.
Hall, B. C. (1965). Subject and object in modern English. PhD Thesis: MIT.
Hirsch, C., & Wexler, K. (2006). Children’s passives and their resulting
interpretation. In K. Deen, J. Nomura, B. Schulz, & B. Schwartz (Eds.), The
proceedings of the inaugural conference on generative approaches to
language acquisition – North America (pp. 125-136). University of
Connecticut occasional papers in linguistics 4.
Hirsh-Pasek, K., Treiman, R., & Schneiderman, M. (1984). Brown & Hanlon
revisited: Mothers’ sensitivity to ungrammatical forms. Journal of Child
Language, 11(1), 81-88.
Horn, J. L. (1965). A rationale and test for the number of factors in factor analysis.
Psychometrika, 30, 179-185.
Huttenlocher, J., Vasilyeva, M., & Shimpi, P. (2004). Syntactic priming in young
children. Journal of Memory and Language, 50(2), 185-195.
Ibbotson, P., & Tomasello, M. (2009). Prototype constructions in early language
acquisition. Language and Cognition, 1(1), 29-85.
185
Israel, M., Johnson, C., & Brooks, P. J. (2000). From states to events: The
acquisition of English passive participles. Cognitive Linguistics, 11(1-2), 103-
129.
Jackendoff, Ray (1977). X-bar-Syntax: A Study of Phrase Structure: Linguistic
Inquiry Monograph 2. Cambridge, MA: MIT Press.
Kerns, G. J. (2007). kmo() function for R. URL
http://tolstoy.newcastle.edu.au/R/e2/help/07/08/22816.html
Kidd, E., Lieven, E., & Tomasello, M. (2006). Examining the role of lexical
frequency in the acquisition and processing of sentential complements.
Cognitive Development, 21(2), 93-107.
Langacker, R. W. (1987). Foundations of cognitive grammar: Theoretical
prerequisites (Vol. 1). Stanford: Stanford University Press.
Langacker, R. W. (2003). A dynamic usage-based model. In M. Barlow & S.
Kemmer (Eds.). Usage-based models of language acquisition, pp.1-63.
Stanford: CSLI.
Lempert, H. (1990). Acquisition of passives: The role of patient, animacy, salience,
and lexical accessibility. Journal of Child Language, 17(3), 677-696.
Levin, B. (1993). English verb classes and alternations: A preliminary investigation.
Chicago: University of Chicago Press.
Levin, B. & Rappaport, M. (1986). The formation of adjectival passives. Linguistic
Inquiry 17(4), 623-661.
Lieven, E., Behrens, H., Speares, J., & Tomasello, M. (2003). Early syntactic
creativity: A usage-based approach. Journal of Child Language, 30(2), 333-
370.
Lieven, E. V. M., Pine, J. M., & Baldwin, G. (1997). Lexically-based learning and
early grammatical development. Journal of Child Language, 24(1), 187-219.
Lieven, E., Salomo, D., & Tomasello, M. (2009). Two-year-old children’s
production of multiword utterances: A usage-based analysis. Cognitive
Linguistics, 20(3), 481-508.
Lord, C. (1979). Don’t you fall me down: Children’s generalizations regarding cause
and transitivity. In Papers and reports on child language development
(PRCLD) 17. Stanford, CA: Stanford University Department of Linguistics.
MacWhinney, B. (1975). Rules, rote, and analogy in morphological formations by
Hungarian children. Journal of Child Language, 2(1), 65-77.
186
MacWhinney, B. (2000) The CHILDES project: Tools for analyzing talk (3rd Ed.).
Mahwah, NJ: Lawrence Erlbaum Associates.
MacWhinney, B. (2004). A multiple process solution to the logical problem of
language acquisition. Journal of Child Language, 31, 883-914.
Maratsos, M., & Chalkey, M. (1980). The internal language of children’s syntax: The
ontogenesis and representation of syntactic categories. In K. Nelson (Ed.),
Children’s language, (Vol. 2, pp. 127-151). New York: Gardner Press.
Maratsos, M., Fox, D., Becker, J., & Chalkey, M. (1985). Semantic restrictions on
children’s passives. Cognition, 19(2), 167-191.
Maratsos, M., Gudeman, R., Gerard-Ngo, P., & DeHart, G. (1987). A study in novel
word learning: The productivity of the causative. In B. MacWhinney (Ed.),
Mechanisms of language acquisition. Hillsdale, NJ: Lawrence Erlbaum
Associates.
Marcus, G. F. (1993). Negative evidence in language acquisition. Cognition, 46, 53-
85.
Matthews, D., & Bannard, C. (2010). Children’s production of unfamiliar word
sequences is predicted by positional variability and latent classes in a large
sample of child-directed speech. Cognitive Science, 34(3), 465-488.
Matthews, D., Lieven, E., Theakston, A., & Tomasello, M. (2005). The role of
frequency in the acquisition of English word order. Cognitive Development,
20(1), 121-136.
Meints, K. (1999). Prototypes and the acquisition of the English passive. In B.
Kokinov (Ed.), Perspectives on cognitive science, Vol. 4, pp.67-77. Sofia:
NBU Press.
Messenger, K., Branigan, H. P., McLean, J. F., & Sorace, A. (2012). Is young
children’s passive syntax semantically constrained? Evidence from syntactic
priming. Journal of Memory and Language, 66(4), 568-587.
Messenger, K., Branigan, H. P., & McLean, J. F. (2011a). Evidence for (shared)
abstract structure underlying children’s short and full passives. Cognition,
121(2), 268-274.
Messenger, K., Branigan, H. P., & McLean, J. F. (2011b). Is children’s acquisition of
the passive a staged process? Evidence from six- and nine-year-olds’
production of passives. Journal of Child Language, 39(5), 991-1016.
187
Morgan, J. L., Bonamo, K. M., & Travis, L. L. (1995). Negative evidence on
negative evidence. Developmental Psychology, 21(2), 180-197.
Morgan, J. L., & Travis, L. L. (1989). Limits on negative information in language
input. Journal of Child Language, 16(3), 531-552.
Naigles, L. R., Fowler. A. & Helm, A. (1992). Developmental Shifts in the
Construction of Verb Meanings. Cognitive Development, 7, 403-427.
Naigles, L. R. & Hoff-Ginsberg, E. (1998). Why are some verbs learned before other
verbs? Effects of input frequency and structure on children’s early verb use.
Journal of Child Language, 25: 95-120.
Naigles, L. R. & Lehrer, N. (2002). Language-general and language-specific
influences on children’s acquisition of argument structure: a comparison on
French and English. Journal of Child Language, 29: 545-566.
Ninio, A. (2014). Variables and values in children’s early word-combinations.
Psychology of Language and Communication, 18(2), 106-125.
Noble, C. H., Rowland, C. F., & Pine, J. M. (2011). Comprehension of argument
structure and semantic roles: Evidence from English-learning children and the
forced-choice pointing paradigm. Cognitive Science, 35(5), 963-982.
Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization
relationship. Journal of Experimental Psychology: General, 115(1), 39-57.
Penner, S. G. (1987). Parental responses to grammatical and ungrammatical child
utterances. Child Development, 58(2), 376-384.
Perfors, A., Tenenbaum, J, B, & Wonnacott, E. (2010). Variability, negative
evidence, and the acquisition of verb argument constructions. Journal of
Child Language, 37: 607-642.
Peter, M., Chang, F., Pine, J.M., Blything, R., & Rowland, C.F. (2015). When and
how do children develop knowledge of verb argument structure? Evidence
from verb bias effects in a structural priming task. Journal of Memory and
Language, 81, 1-15.
Pickering, M. J., & Ferreira, V. S. (2008). Structural priming: A critical review.
Psychological Bulletin, 134(3), 427-459.
Pine, J. M., & Lieven, E. V. M. (1993). Reanalysing rote-leaned phrases: Individual
differences in the transition to multi-word speech. Journal of Child
Language, 20(3), 551-571.
188
Pine, J. M., Lieven, E. V. M., & Rowland, C. F. (1998). Comparing different models
of the development of the English verb category. Linguistics, 36(4), 807-830.
Pinker, S. (1984). Language learnability and language development. Cambridge,
MA: Harvard University Press.
Pinker, S. (1989). Learnability and cognition: The acquisition of argument structure.
Cambridge. MA: MIT.
Pinker, S., Lebeaux, D. S., & Frost, L. A. (1987). Productivity and constraints in the
acquisition of the passive. Cognition, 26(3), 195-267.
Pollard, C., & Sag, I. A. (1994). Head-Driven Phrase Structure Grammar. Chicago:
University of Chicago Press.
Pye, C. (1990). The acquisition of ergative languages. Linguistics, 28(6), 1291-1330.
R Core Team (2014). R: A language and environment for statistical computing.
Vienna: R Foundation for Statistical Computing. http://www.r-project.org/.
R Core Team (2015). R: A language and environment for statistical computing. R
Foundation for Statistical Computing, Vienna, Austria. http://www.R-
project.org.
Reber, A. S. (1969) Transfer of syntactic structure in synthetic languages. Journals
of Experimental Psychology, 81(1), 115-119.
Reber, A. S. (1989). Implicit learning and tacit knowledge. Journal of Experimental
Psychology: General, 118(3), 219-235.
Revelle, W. (2014). psych: Procedures for personality and psychological research.
Version 1.4.6. Evanston, IL: Northwestern University. http://cran.r-
project.org/package=psych.
Rosch, E. (1975). Cognitive representations of semantic categories. Journal of
experimental psychology: General, 104(3), 192-233.
Rowland, C. F., Chang, F., Ambridge, B., Pine, J. M., & Lieven, E. V. M. (2012).
The development of abstract syntax: Evidence from structural priming and
the lexical boost. Cognition, 125(1), 49–63.
Rowland, C. F., & Theakston, A. L. (2009). The acquisition of auxiliary syntax: A
longitudinal elicitation study. Part 2: The modals and auxiliary DO. Journal
of Speech, Language and Hearing Research, 52(6), 1471-1492.
189
Savage, C., Lieven, E. V. M., Theakston, A., & Tomasello, M. (2003). Testing the
abstractness of children’s linguistic representations: Lexical and structural
priming of syntactic constructions in young children. Developmental Science,
6(5), 557-567.
Savage, C., Lieven, E. V. M., Theakston, A., & Tomasello, M. (2006). Structural
priming as implicit learning in language acquisition: The persistence of
lexical and structural priming in 4-year-olds. Language Learning and
Development, 2(1), 27-149.
Saxton, M., Backley, P. & Gallaway, C. (2005). Negative input for grammatical
errors: Effects after a lag of 12 weeks. Journal of Child Language, 32(3),
643-672.
Saxton, M., Kulcsar, B., Marshall, G., & Rupra, M. (1998). Longer-term effects of
corrective input: An experimental approach. Journal of Child Language,
25(3), 701-721.
Schäfer, J., Opgen-Rhein, R., Zuber, V., Ahdesmäki, M., Duarte Silva, A.P., &
Strimmer, K. (2013). corpcor: Efficient estimation of covariance and (partial)
correlation. Version 1.6.6. http://cran.r-project.org/package=corpcor.
Schlesinger, I. M. (1988). The origin of relational categories. In Y. Levy, I. M.
Schlesinger & M. D. S. Braine (Eds.), Categories and processes in language
acquisition (pp. 121-178). Hillsdale, NJ: Lawrence Erlbaum Associates, Inc.
Stefanowitsch, A. (2008). Negative evidence and preemption: A constructional
approach to ungrammaticality. Cognitive Linguistics, 19(3), 513-531.
Stevens, J. P. (2012). Applied multivariate statistics for the social sciences (5th Ed.).
New York: Routledge.
Street, J. A., & Dąbrowska, E. (2014). Lexically specific knowledge and individual
differences in adult native speakers’ processing of the English passive.
Applied Psycholonguistics, 31(1), 97-118.
Sudhalter, V. & Braine, M. D. S. (1985). How does comprehension of passives
develop? A comparison of actional and experiential verbs. Journal of Child
Language, 12(2), 455-470.
Theakston, A. L. (2004). The role of entrenchment in children's and adults'
performance on grammaticality judgement tasks. Cognitive Development,
19(1), 15-34.
190
Theakston, A. L., Lieven, E. V. M., Pine, J. M., & Rowland, C. F. (2005) The
acquisition of auxiliary syntax: BE and HAVE. Cognitive Linguistics, 16(1),
247-277.
Thothathiri, M. & Snedeker, J. (2008) Syntactic priming during language
comprehension in three- and four-year-old children. Journal of Memory and
Language, 58(2), 188-213.
Tomasello, M. (1992). First verbs: A case study of grammatical development.
Cambridge: Cambridge University Press.
Tomasello, M. (2003). Constructing a language: A usage-based theory of language
acquisition. Cambridge, MA: Harvard University Press.
Tomasello, M., Brooks, P. J., & Stern, E. (1998). Learning to produce passive
utterances through discourse. First Language, 18(53), 223-237.
Tversky, A., & Kahneman, D. (1971). Belief in the law of small numbers.
Psychological Bulletin, 76(2), 105-110.
Valian, V., Solt, S., & Stewert, J. (2009). Abstract categories or limited-scope
formulae? The case of children’s determiners. Journal of Child Language,
36(4), 743-778.
van Valin, R. D. (2005). Exploring the syntax-semantics interface. Cambridge, UK:
Cambridge University Press.
Wexler, K. (1998). Very early parameter setting and the unique checking constraint:
A new explanation of the optional infinitive stage. Lingua, 106(1), 23-79.
Wilson, S. (2003). Lexically specific construction in the acquisition of inflection in
English. Journal of Child Language, 30(1), 75-115.
Wonnacott, E. (2011). Balancing generalization and lexical conservatism: An
artificial language study with child learners. Journal of Memory and
Language, 65, 1-14.
